Go to Qiita Advent Calendar 2024 Top
LoginSignup
17
11

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 5 years have passed since last update.

@PINTO

Learn "Openpose" from scratch with MobileNetv2 + MS-COCO and deploy it to OpenVINO/TensorflowLite (Inference by OpenVINO/NCS2) Part.2

Last updated at Posted at 2019-05-03

MobileNetV2-PoseEstimation　GitHub stars

Tensorflow-bin　GitHub stars

１．Introduction

前回記事 Learn "Openpose" from scratch with MobileNetv2 + MS-COCO and deploy it to OpenVINO/TensorflowLite Part.1 により、Openpose のトレーニング実施と OpenVINO / Tensorflow Lite モデルを生成しました。
今回は作成した学習済みファイルとPythonでOpenVINOとTensorflow Liteのリアルタイム骨格検出を行います。
Tensorflow-GPU および CUDA を使用せず、 PythonとOpenVINO で、 CPUのみ による推論と描画を行います。

何故私がここまでCPU推論にこだわっているかというと、電気を大食いして高熱を発するGPUや、設置スペースを無駄に消費する外付けデバイス、インタフェースの性能が低いことに引きずられて十分なパフォーマンスが得られないデバイス、などは実戦投入するときに扱いづらいと考えているからです。 また、手軽さとは裏腹に、 連続稼働数時間で停止してしまうような不安定なインフラはとても良くない です。

余談はさておき、GPUを使用せずにどれほどの性能が得られるでしょうか。
前回記事投稿時のスピードは下図のとおり、超遅いです。

【前回】 Tensorflow-CPU's Test Core i7, FP32 (disabled OpenVINO/Tensorflow Lite) - 4 FPS
Youtube： https://youtu.be/nEKc7VIm42A
ezgif.com-optimize (28).gif

では、いつもどおりいきなり結果に行きます。 前回と同じモデルを使用し、OpenVINO対応しただけです。
前回・今回ともにGPUを使用していません。 今回は精度とスピードのバランスを計るため、２種類のモデルを生成して動作検証しました。
今回はCPU推論にも関わらずフレームレートが１０倍ほど向上しています。
動画はUSBカメラの撮影レートを30FPSに制限していますが、高速なフレームレートで撮影すると40FPSまで速度が向上します。 （つまり、30FPSのUSBカメラの撮影スピードより推論スピードのほうが速いという結果です。）
ちなみに、 Core i7 でも Core m3 でも結果は同じです。 また、Multi Person に対応していますので、結構負荷を掛けています。

ノーマルモードはmobilenet_v2_1.4_224、ブーストモードはmobilenet_v2_0.5_224を使用して学習しています。

【今回】 Boost Mode, OpenVINO Test Core i7 CPU only (disabled GPU), Sync - FP32 - 40 FPS++
Youtube： https://youtu.be/J-a2kHS4nTc
03.gif

【今回】 Boost Mode, OpenVINO Test NCS2 x1, Async - FP16 - 25 FPS
Youtube： https://youtu.be/CUAojvJYRLE
03.gif

【今回】 Normal Mode, OpenVINO Test Core i7 CPU only, Sync - FP32 (enabled OpenVINO) - 15-19 FPS
Youtube： https://youtu.be/t66bnWwZ7rE
01.gif

【今回】 Normal Mode, OpenVINO Test NCS2 x1, Async - FP16 - 8-10 FPS
Youtube： https://youtu.be/hFFUectGQ2A
ezgif.com-optimize (30).gif

２．Environment

３．Procedure

手順もなにも無く、この記事の一番上にあるニコちゃんマークのリンク先のリポジトリをクローンして実行するだけです。
一応、作成したロジック群を全て下記に転記しておきます。 みなさんのご参考になれば幸いです。
非同期モードのプログラムは NCS2 の複数本挿しによるブーストにも対応しています。
なお、Tensorflow Lite版は遅すぎて使い物になりませんでしたので、ロジックのご紹介だけにとどめておきます。

openvino-usbcamera-cpu-ncs2-sync.py_同期推論モード
import sys
import os
import numpy as np
import cv2
from os import system
import io
import time
from os.path import isfile
from os.path import join
import re
import argparse
import platform
try:
    from armv7l.openvino.inference_engine import IENetwork, IEPlugin
except:
    from openvino.inference_engine import IENetwork, IEPlugin


def getKeypoints(probMap, threshold=0.1):

    mapSmooth = cv2.GaussianBlur(probMap, (3, 3), 0, 0)
    mapMask = np.uint8(mapSmooth>threshold)
    keypoints = []
    contours = None
    try:
        #OpenCV4.x
        contours, _ = cv2.findContours(mapMask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    except:
        #OpenCV3.x
        _, contours, _ = cv2.findContours(mapMask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

    for cnt in contours:
        blobMask = np.zeros(mapMask.shape)
        blobMask = cv2.fillConvexPoly(blobMask, cnt, 1)
        maskedProbMap = mapSmooth * blobMask
        _, maxVal, _, maxLoc = cv2.minMaxLoc(maskedProbMap)
        keypoints.append(maxLoc + (probMap[maxLoc[1], maxLoc[0]],))

    return keypoints


def getValidPairs(outputs, w, h):
    valid_pairs = []
    invalid_pairs = []
    n_interp_samples = 10
    paf_score_th = 0.1
    conf_th = 0.7

    for k in range(len(mapIdx)):
        pafA = outputs[0, mapIdx[k][0], :, :]
        pafB = outputs[0, mapIdx[k][1], :, :]
        pafA = cv2.resize(pafA, (w, h))
        pafB = cv2.resize(pafB, (w, h))

        candA = detected_keypoints[POSE_PAIRS[k][0]]
        candB = detected_keypoints[POSE_PAIRS[k][1]]
        nA = len(candA)
        nB = len(candB)

        if( nA != 0 and nB != 0):
            valid_pair = np.zeros((0,3))
            for i in range(nA):
                max_j=-1
                maxScore = -1
                found = 0
                for j in range(nB):
                    d_ij = np.subtract(candB[j][:2], candA[i][:2])
                    norm = np.linalg.norm(d_ij)
                    if norm:
                        d_ij = d_ij / norm
                    else:
                        continue
                    interp_coord = list(zip(np.linspace(candA[i][0], candB[j][0], num=n_interp_samples),
                                            np.linspace(candA[i][1], candB[j][1], num=n_interp_samples)))
                    paf_interp = []
                    for k in range(len(interp_coord)):
                        paf_interp.append([pafA[int(round(interp_coord[k][1])), int(round(interp_coord[k][0]))],
                                           pafB[int(round(interp_coord[k][1])), int(round(interp_coord[k][0]))] ])
                    paf_scores = np.dot(paf_interp, d_ij)
                    avg_paf_score = sum(paf_scores)/len(paf_scores)

                    if ( len(np.where(paf_scores > paf_score_th)[0]) / n_interp_samples ) > conf_th :
                        if avg_paf_score > maxScore:
                            max_j = j
                            maxScore = avg_paf_score
                            found = 1
                if found:
                    valid_pair = np.append(valid_pair, [[candA[i][3], candB[max_j][3], maxScore]], axis=0)

            valid_pairs.append(valid_pair)
        else:
            invalid_pairs.append(k)
            valid_pairs.append([])
    return valid_pairs, invalid_pairs


def getPersonwiseKeypoints(valid_pairs, invalid_pairs):
    personwiseKeypoints = -1 * np.ones((0, 19))

    for k in range(len(mapIdx)):
        if k not in invalid_pairs:
            partAs = valid_pairs[k][:,0]
            partBs = valid_pairs[k][:,1]
            indexA, indexB = np.array(POSE_PAIRS[k])

            for i in range(len(valid_pairs[k])):
                found = 0
                person_idx = -1
                for j in range(len(personwiseKeypoints)):
                    if personwiseKeypoints[j][indexA] == partAs[i]:
                        person_idx = j
                        found = 1
                        break

                if found:
                    personwiseKeypoints[person_idx][indexB] = partBs[i]
                    personwiseKeypoints[person_idx][-1] += keypoints_list[partBs[i].astype(int), 2] + valid_pairs[k][i][2]

                elif not found and k < 17:
                    row = -1 * np.ones(19)
                    row[indexA] = partAs[i]
                    row[indexB] = partBs[i]
                    row[-1] = sum(keypoints_list[valid_pairs[k][i,:2].astype(int), 2]) + valid_pairs[k][i][2]
                    personwiseKeypoints = np.vstack([personwiseKeypoints, row])
    return personwiseKeypoints


fps = ""
detectfps = ""
framecount = 0
time1 = 0

camera_width = 320
camera_height = 240

keypointsMapping = ['Nose', 'Neck', 'R-Sho', 'R-Elb', 'R-Wr', 'L-Sho', 'L-Elb', 'L-Wr', 'R-Hip', 'R-Knee', 'R-Ank', 'L-Hip', 'L-Knee', 'L-Ank', 'R-Eye', 'L-Eye', 'R-Ear', 'L-Ear']
POSE_PAIRS = [[1,2], [1,5], [2,3], [3,4], [5,6], [6,7], [1,8], [8,9], [9,10], [1,11], [11,12], [12,13], [1,0], [0,14], [14,16], [0,15], [15,17], [2,17], [5,16]]
mapIdx = [[31,32], [39,40], [33,34], [35,36], [41,42], [43,44], [19,20], [21,22], [23,24], [25,26], [27,28], [29,30], [47,48], [49,50], [53,54], [51,52], [55,56], [37,38], [45,46]]
colors = [[0,100,255], [0,100,255], [0,255,255], [0,100,255], [0,255,255], [0,100,255], [0,255,0], [255,200,100], [255,0,255], [0,255,0], [255,200,100], [255,0,255], [0,0,255], [255,0,0], [200,200,0], [255,0,0], [200,200,0], [0,0,0]]

cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FPS, 30)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, camera_width)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, camera_height)

parser = argparse.ArgumentParser()
parser.add_argument("-d", "--device", help="Specify the target device to infer on; CPU, GPU, MYRIAD is acceptable. (Default=CPU)", default="CPU", type=str)
parser.add_argument("-b", "--boost", help="Setting it to True will make it run faster instead of sacrificing accuracy. (Default=False)", default=False, type=bool)
args = parser.parse_args()

plugin = IEPlugin(device=args.device)

if "CPU" == args.device:
    if platform.processor() == "x86_64":
        plugin.add_cpu_extension("lib/libcpu_extension.so")
    if args.boost == False:
        model_xml = "models/train/test/openvino/mobilenet_v2_1.4_224/FP32/frozen-model.xml"
    else:
        model_xml = "models/train/test/openvino/mobilenet_v2_0.5_224/FP32/frozen-model.xml"

elif "GPU" == args.device or "MYRIAD" == args.device:
    if args.boost == False:
        model_xml = "models/train/test/openvino/mobilenet_v2_1.4_224/FP16/frozen-model.xml"
    else:
        model_xml = "models/train/test/openvino/mobilenet_v2_0.5_224/FP16/frozen-model.xml"

else:
    print("Specify the target device to infer on; CPU, GPU, MYRIAD is acceptable.")
    sys.exit(0)

model_bin = os.path.splitext(model_xml)[0] + ".bin"
net = IENetwork(model=model_xml, weights=model_bin)
input_blob = next(iter(net.inputs))
exec_net = plugin.load(network=net)
inputs = net.inputs["image"]

h = inputs.shape[2] #368
w = inputs.shape[3] #432

threshold = 0.1
nPoints = 18

try:

    while True:
        t1 = time.perf_counter()

        ret, color_image = cap.read()
        if not ret:
            break

        colw = color_image.shape[1]
        colh = color_image.shape[0]
        new_w = int(colw * min(w/colw, h/colh))
        new_h = int(colh * min(w/colw, h/colh))

        resized_image = cv2.resize(color_image, (new_w, new_h), interpolation = cv2.INTER_CUBIC)
        canvas = np.full((h, w, 3), 128)
        canvas[(h - new_h)//2:(h - new_h)//2 + new_h,(w - new_w)//2:(w - new_w)//2 + new_w, :] = resized_image

        prepimg = canvas
        prepimg = prepimg[np.newaxis, :, :, :]     # Batch size axis add
        prepimg = prepimg.transpose((0, 3, 1, 2))  # NHWC to NCHW, (1, 3, 368, 432)
        outputs = exec_net.infer(inputs={input_blob: prepimg})["Openpose/concat_stage7"]

        detected_keypoints = []
        keypoints_list = np.zeros((0, 3))
        keypoint_id = 0

        for part in range(nPoints):
            probMap = outputs[0, part, :, :]
            probMap = cv2.resize(probMap, (canvas.shape[1], canvas.shape[0])) # (432, 368)
            keypoints = getKeypoints(probMap, threshold)
            keypoints_with_id = []

            for i in range(len(keypoints)):
                keypoints_with_id.append(keypoints[i] + (keypoint_id,))
                keypoints_list = np.vstack([keypoints_list, keypoints[i]])
                keypoint_id += 1

            detected_keypoints.append(keypoints_with_id)

        frameClone = np.uint8(canvas.copy())
        for i in range(nPoints):
            for j in range(len(detected_keypoints[i])):
                cv2.circle(frameClone, detected_keypoints[i][j][0:2], 5, colors[i], -1, cv2.LINE_AA)

        valid_pairs, invalid_pairs = getValidPairs(outputs, w, h)
        personwiseKeypoints = getPersonwiseKeypoints(valid_pairs, invalid_pairs)

        for i in range(17):
            for n in range(len(personwiseKeypoints)):
                index = personwiseKeypoints[n][np.array(POSE_PAIRS[i])]
                if -1 in index:
                    continue
                B = np.int32(keypoints_list[index.astype(int), 0])
                A = np.int32(keypoints_list[index.astype(int), 1])
                cv2.line(frameClone, (B[0], A[0]), (B[1], A[1]), colors[i], 3, cv2.LINE_AA)

        cv2.putText(frameClone, fps, (w-170,15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)

        cv2.namedWindow("USB Camera", cv2.WINDOW_AUTOSIZE)
        cv2.imshow("USB Camera" , frameClone)

        if cv2.waitKey(1)&0xFF == ord('q'):
            break

        # FPS calculation
        framecount += 1
        if framecount >= 15:
            fps = "(Playback) {:.1f} FPS".format(time1/15)
            framecount = 0
            time1 = 0
        t2 = time.perf_counter()
        elapsedTime = t2-t1
        time1 += 1/elapsedTime

except:
    import traceback
    traceback.print_exc()

finally:

    print("\n\nFinished\n\n")
openvino-usbcamera-cpu-ncs2-async.py_非同期推論モード
import sys
import os
import numpy as np
import cv2
from os import system
import io
import time
from os.path import isfile
from os.path import join
import re
import argparse
import platform
try:
    from armv7l.openvino.inference_engine import IENetwork, IEPlugin
except:
    from openvino.inference_engine import IENetwork, IEPlugin
import multiprocessing as mp
from time import sleep
import threading
import heapq


def getKeypoints(probMap, threshold=0.1):

    mapSmooth = cv2.GaussianBlur(probMap, (3, 3), 0, 0)
    mapMask = np.uint8(mapSmooth>threshold)
    keypoints = []
    contours = None
    try:
        #OpenCV4.x
        contours, _ = cv2.findContours(mapMask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    except:
        #OpenCV3.x
        _, contours, _ = cv2.findContours(mapMask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

    for cnt in contours:
        blobMask = np.zeros(mapMask.shape)
        blobMask = cv2.fillConvexPoly(blobMask, cnt, 1)
        maskedProbMap = mapSmooth * blobMask
        _, maxVal, _, maxLoc = cv2.minMaxLoc(maskedProbMap)
        keypoints.append(maxLoc + (probMap[maxLoc[1], maxLoc[0]],))

    return keypoints


def getValidPairs(detected_keypoints, outputs, w, h):
    valid_pairs = []
    invalid_pairs = []
    n_interp_samples = 10
    paf_score_th = 0.1
    conf_th = 0.7

    for k in range(len(mapIdx)):
        pafA = outputs[0, mapIdx[k][0], :, :]
        pafB = outputs[0, mapIdx[k][1], :, :]
        pafA = cv2.resize(pafA, (w, h))
        pafB = cv2.resize(pafB, (w, h))

        candA = detected_keypoints[POSE_PAIRS[k][0]]
        candB = detected_keypoints[POSE_PAIRS[k][1]]
        nA = len(candA)
        nB = len(candB)

        if( nA != 0 and nB != 0):
            valid_pair = np.zeros((0,3))
            for i in range(nA):
                max_j=-1
                maxScore = -1
                found = 0
                for j in range(nB):
                    d_ij = np.subtract(candB[j][:2], candA[i][:2])
                    norm = np.linalg.norm(d_ij)
                    if norm:
                        d_ij = d_ij / norm
                    else:
                        continue
                    interp_coord = list(zip(np.linspace(candA[i][0], candB[j][0], num=n_interp_samples),
                                            np.linspace(candA[i][1], candB[j][1], num=n_interp_samples)))
                    paf_interp = []
                    for k in range(len(interp_coord)):
                        paf_interp.append([pafA[int(round(interp_coord[k][1])), int(round(interp_coord[k][0]))],
                                           pafB[int(round(interp_coord[k][1])), int(round(interp_coord[k][0]))] ])
                    paf_scores = np.dot(paf_interp, d_ij)
                    avg_paf_score = sum(paf_scores)/len(paf_scores)

                    if ( len(np.where(paf_scores > paf_score_th)[0]) / n_interp_samples ) > conf_th :
                        if avg_paf_score > maxScore:
                            max_j = j
                            maxScore = avg_paf_score
                            found = 1
                if found:
                    valid_pair = np.append(valid_pair, [[candA[i][3], candB[max_j][3], maxScore]], axis=0)

            valid_pairs.append(valid_pair)
        else:
            invalid_pairs.append(k)
            valid_pairs.append([])
    return valid_pairs, invalid_pairs


def getPersonwiseKeypoints(valid_pairs, invalid_pairs, keypoints_list):
    personwiseKeypoints = -1 * np.ones((0, 19))

    for k in range(len(mapIdx)):
        if k not in invalid_pairs:
            partAs = valid_pairs[k][:,0]
            partBs = valid_pairs[k][:,1]
            indexA, indexB = np.array(POSE_PAIRS[k])

            for i in range(len(valid_pairs[k])):
                found = 0
                person_idx = -1
                for j in range(len(personwiseKeypoints)):
                    if personwiseKeypoints[j][indexA] == partAs[i]:
                        person_idx = j
                        found = 1
                        break

                if found:
                    personwiseKeypoints[person_idx][indexB] = partBs[i]
                    personwiseKeypoints[person_idx][-1] += keypoints_list[partBs[i].astype(int), 2] + valid_pairs[k][i][2]

                elif not found and k < 17:
                    row = -1 * np.ones(19)
                    row[indexA] = partAs[i]
                    row[indexB] = partBs[i]
                    row[-1] = sum(keypoints_list[valid_pairs[k][i,:2].astype(int), 2]) + valid_pairs[k][i][2]
                    personwiseKeypoints = np.vstack([personwiseKeypoints, row])
    return personwiseKeypoints


processes = []
fps = ""
detectfps = ""
framecount = 0
detectframecount = 0
time1 = 0
time2 = 0
lastresults = None

keypointsMapping = ['Nose', 'Neck', 'R-Sho', 'R-Elb', 'R-Wr', 'L-Sho', 'L-Elb', 'L-Wr', 'R-Hip', 'R-Knee', 'R-Ank', 'L-Hip', 'L-Knee', 'L-Ank', 'R-Eye', 'L-Eye', 'R-Ear', 'L-Ear']
POSE_PAIRS = [[1,2], [1,5], [2,3], [3,4], [5,6], [6,7], [1,8], [8,9], [9,10], [1,11], [11,12], [12,13], [1,0], [0,14], [14,16], [0,15], [15,17], [2,17], [5,16]]
mapIdx = [[31,32], [39,40], [33,34], [35,36], [41,42], [43,44], [19,20], [21,22], [23,24], [25,26], [27,28], [29,30], [47,48], [49,50], [53,54], [51,52], [55,56], [37,38], [45,46]]
colors = [[0,100,255], [0,100,255], [0,255,255], [0,100,255], [0,255,255], [0,100,255], [0,255,0], [255,200,100], [255,0,255], [0,255,0], [255,200,100], [255,0,255], [0,0,255], [255,0,0], [200,200,0], [255,0,0], [200,200,0], [0,0,0]]


def image_preprocessing(color_image, w, h, new_w, new_h):
    resized_image = cv2.resize(color_image, (new_w, new_h), interpolation = cv2.INTER_CUBIC)
    canvas = np.full((h, w, 3), 128)
    canvas[(h-new_h)//2:(h-new_h)//2 + new_h,(w-new_w)//2:(w-new_w)//2 + new_w,  :] = resized_image
    return canvas


def camThread(results, frameBuffer, camera_width, camera_height, vidfps, nPoints, w, h, new_w, new_h):
    global fps
    global detectfps
    global lastresults
    global framecount
    global detectframecount
    global time1
    global time2
    global cam
    global window_name

    cam = cv2.VideoCapture(0)
    if cam.isOpened() != True:
        print("USB Camera Open Error!!!")
        sys.exit(0)
    cam.set(cv2.CAP_PROP_FPS, vidfps)
    cam.set(cv2.CAP_PROP_FRAME_WIDTH, camera_width)
    cam.set(cv2.CAP_PROP_FRAME_HEIGHT, camera_height)
    window_name = "USB Camera"
    wait_key_time = 1

    cv2.namedWindow(window_name, cv2.WINDOW_AUTOSIZE)

    while True:
        t1 = time.perf_counter()

        # USB Camera Stream Read
        s, color_image = cam.read()
        if not s:
            continue
        if frameBuffer.full():
            frameBuffer.get()

        color_image = image_preprocessing(color_image.copy(), w, h, new_w, new_h)
        frameClone = np.uint8(color_image.copy())
        frameBuffer.put(color_image)

        if not results.empty():
            detected_keypoints, outputs, keypoints_list = results.get(False)
            detectframecount += 1

            for i in range(nPoints):
                for j in range(len(detected_keypoints[i])):
                    cv2.circle(frameClone, detected_keypoints[i][j][0:2], 5, colors[i], -1, cv2.LINE_AA)

            valid_pairs, invalid_pairs = getValidPairs(detected_keypoints, outputs, w, h)
            personwiseKeypoints = getPersonwiseKeypoints(valid_pairs, invalid_pairs, keypoints_list)

            for i in range(17):
                for n in range(len(personwiseKeypoints)):
                    index = personwiseKeypoints[n][np.array(POSE_PAIRS[i])]
                    if -1 in index:
                        continue
                    B = np.int32(keypoints_list[index.astype(int), 0])
                    A = np.int32(keypoints_list[index.astype(int), 1])
                    cv2.line(frameClone, (B[0], A[0]), (B[1], A[1]), colors[i], 3, cv2.LINE_AA)

            cv2.putText(frameClone, fps, (w-170,15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)
            lastresults = [detected_keypoints, outputs, keypoints_list]
        else:
            if not isinstance(lastresults, type(None)):
                detected_keypoints, outputs, keypoints_list = lastresults
                for i in range(nPoints):
                    for j in range(len(detected_keypoints[i])):
                        cv2.circle(frameClone, detected_keypoints[i][j][0:2], 5, colors[i], -1, cv2.LINE_AA)

                valid_pairs, invalid_pairs = getValidPairs(detected_keypoints, outputs, w, h)
                personwiseKeypoints = getPersonwiseKeypoints(valid_pairs, invalid_pairs, keypoints_list)

                for i in range(17):
                    for n in range(len(personwiseKeypoints)):
                        index = personwiseKeypoints[n][np.array(POSE_PAIRS[i])]
                        if -1 in index:
                            continue
                        B = np.int32(keypoints_list[index.astype(int), 0])
                        A = np.int32(keypoints_list[index.astype(int), 1])
                        cv2.line(frameClone, (B[0], A[0]), (B[1], A[1]), colors[i], 3, cv2.LINE_AA)

        cv2.putText(frameClone, fps,       (w-170,15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)
        cv2.putText(frameClone, detectfps, (w-170,30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)
        cv2.imshow(window_name, frameClone)

        if cv2.waitKey(wait_key_time)&0xFF == ord('q'):
            sys.exit(0)

        ## Print FPS
        framecount += 1
        if framecount >= 15:
            fps       = "(Playback) {:.1f} FPS".format(time1/15)
            detectfps = "(Detection) {:.1f} FPS".format(detectframecount/time2)
            framecount = 0
            detectframecount = 0
            time1 = 0
            time2 = 0
        t2 = time.perf_counter()
        elapsedTime = t2-t1
        time1 += 1/elapsedTime
        time2 += elapsedTime


# l = Search list
# x = Search target value
def searchlist(l, x, notfoundvalue=-1):
    if x in l:
        return l.index(x)
    else:
        return notfoundvalue


def async_infer(ncsworker):

    #ncsworker.skip_frame_measurement()

    while True:
        ncsworker.predict_async()


class NcsWorker(object):

    def __init__(self, devid, device, model_xml, frameBuffer, results, camera_width, camera_height, number_of_ncs, vidfps, nPoints, w, h, new_w, new_h):
        self.devid = devid
        self.frameBuffer = frameBuffer
        self.model_xml = model_xml
        self.model_bin = os.path.splitext(model_xml)[0] + ".bin"
        self.camera_width = camera_width
        self.camera_height = camera_height
        self.threshold = 0.1
        self.nPoints = nPoints
        self.num_requests = 4
        self.inferred_request = [0] * self.num_requests
        self.heap_request = []
        self.inferred_cnt = 0
        self.plugin = IEPlugin(device=device)
        if "CPU" == device:
            if platform.processor() == "x86_64":
                self.plugin.add_cpu_extension("lib/libcpu_extension.so")
        self.net = IENetwork(model=self.model_xml, weights=self.model_bin)
        self.input_blob = next(iter(self.net.inputs))
        self.exec_net = self.plugin.load(network=self.net, num_requests=self.num_requests)
        self.results = results
        self.number_of_ncs = number_of_ncs
        self.predict_async_time = 250
        self.skip_frame = 0
        self.roop_frame = 0
        self.vidfps = vidfps
        self.w = w #432
        self.h = h #368
        self.new_w = new_w
        self.new_h = new_h


    def skip_frame_measurement(self):
            surplustime_per_second = (1000 - self.predict_async_time)
            if surplustime_per_second > 0.0:
                frame_per_millisecond = (1000 / self.vidfps)
                total_skip_frame = surplustime_per_second / frame_per_millisecond
                self.skip_frame = int(total_skip_frame / self.num_requests)
            else:
                self.skip_frame = 0


    def predict_async(self):
        try:

            if self.frameBuffer.empty():
                return

            self.roop_frame += 1
            if self.roop_frame <= self.skip_frame:
               self.frameBuffer.get()
               return
            self.roop_frame = 0

            prepimg = self.frameBuffer.get()
            reqnum = searchlist(self.inferred_request, 0)

            if reqnum > -1:
                prepimg = prepimg[np.newaxis, :, :, :]     # Batch size axis add
                prepimg = prepimg.transpose((0, 3, 1, 2))  # NHWC to NCHW, (1, 3, 368, 432)
                self.exec_net.start_async(request_id=reqnum, inputs={self.input_blob: prepimg})
                self.inferred_request[reqnum] = 1
                self.inferred_cnt += 1
                if self.inferred_cnt == sys.maxsize:
                    self.inferred_request = [0] * self.num_requests
                    self.heap_request = []
                    self.inferred_cnt = 0
                heapq.heappush(self.heap_request, (self.inferred_cnt, reqnum))

            try:
                cnt, dev = heapq.heappop(self.heap_request)
            except:
                return

            if self.exec_net.requests[dev].wait(0) == 0:
                self.exec_net.requests[dev].wait(-1)

                detected_keypoints = []
                keypoints_list = np.zeros((0, 3))
                keypoint_id = 0

                outputs = self.exec_net.requests[dev].outputs["Openpose/concat_stage7"]
                for part in range(self.nPoints):
                    probMap = outputs[0, part, :, :]
                    probMap = cv2.resize(probMap, (self.w, self.h)) # (432, 368)
                    keypoints = getKeypoints(probMap, self.threshold)
                    keypoints_with_id = []

                    for i in range(len(keypoints)):
                        keypoints_with_id.append(keypoints[i] + (keypoint_id,))
                        keypoints_list = np.vstack([keypoints_list, keypoints[i]])
                        keypoint_id += 1

                    detected_keypoints.append(keypoints_with_id)

                self.results.put([detected_keypoints, outputs, keypoints_list])
                self.inferred_request[dev] = 0
            else:
                heapq.heappush(self.heap_request, (cnt, dev))
        except:
            import traceback
            traceback.print_exc()


def inferencer(device, model_xml, results, frameBuffer, number_of_ncs, camera_width, camera_height, vidfps, nPoints, w, h, new_w, new_h):

    # Init infer threads
    threads = []
    for devid in range(number_of_ncs):
        thworker = threading.Thread(target=async_infer, args=(NcsWorker(devid, device, model_xml, frameBuffer, results, camera_width, camera_height, number_of_ncs, vidfps, nPoints, w, h, new_w, new_h),))
        thworker.start()
        threads.append(thworker)

    for th in threads:
        th.join()


if __name__ == '__main__':

    parser = argparse.ArgumentParser()
    parser.add_argument("-d", "--device", help="Specify the target device to infer on; CPU, GPU, MYRIAD is acceptable. (Default=CPU)", default="CPU", type=str)
    parser.add_argument('-numncs','--numberofncs',dest='number_of_ncs',type=int,default=1,help='Number of NCS. (Default=1)')
    parser.add_argument("-b", "--boost", help="Setting it to True will make it run faster instead of sacrificing accuracy. (Default=False)", default=False, type=bool)
    args = parser.parse_args()

    device = args.device

    if "CPU" == device:
        number_of_ncs = 1
        if args.boost == False:
            model_xml = "models/train/test/openvino/mobilenet_v2_1.4_224/FP32/frozen-model.xml"
        else:
            model_xml = "models/train/test/openvino/mobilenet_v2_0.5_224/FP32/frozen-model.xml"

    elif "MYRIAD" == device:
        number_of_ncs = args.number_of_ncs
        if args.boost == False:
            model_xml = "models/train/test/openvino/mobilenet_v2_1.4_224/FP16/frozen-model.xml"
        else:
            model_xml = "models/train/test/openvino/mobilenet_v2_0.5_224/FP16/frozen-model.xml"

    elif "GPU" == device:
        number_of_ncs = 1
        if args.boost == False:
            model_xml = "models/train/test/openvino/mobilenet_v2_1.4_224/FP16/frozen-model.xml"
        else:
            model_xml = "models/train/test/openvino/mobilenet_v2_0.5_224/FP16/frozen-model.xml"

    else:
        print("Specify the target device to infer on; CPU, GPU, MYRIAD is acceptable.")
        sys.exit(0)

    camera_width = 320
    camera_height = 240
    vidfps = 30
    nPoints = 18
    w = 432 # Network size (Width)
    h = 368 # Network size (Height)
    new_w = int(camera_width * min(w/camera_width, h/camera_height))
    new_h = int(camera_height * min(w/camera_width, h/camera_height))

    try:

        mp.set_start_method('forkserver')
        frameBuffer = mp.Queue(4)
        results = mp.Queue()

        # Start detection MultiStick
        # Activation of inferencer
        p = mp.Process(target=inferencer, args=(device, model_xml, results, frameBuffer, number_of_ncs, camera_width, camera_height, vidfps, nPoints, w, h, new_w, new_h), daemon=True)
        p.start()
        processes.append(p)

        if device == "MYRIAD":
            sleep(number_of_ncs * 7)

        # Start streaming
        p = mp.Process(target=camThread, args=(results, frameBuffer, camera_width, camera_height, vidfps, nPoints, w, h, new_w, new_h), daemon=True)
        p.start()
        processes.append(p)

        while True:
            sleep(1)

    except:
        import traceback
        traceback.print_exc()
    finally:
        for p in range(len(processes)):
            processes[p].terminate()

        print("\n\nFinished\n\n")
tflite_usbcamera_cpu-sync.py_TensorflowLite版
import cv2, sys, time
import numpy as np
import tensorflow as tf
from PIL import Image


def getKeypoints(probMap, threshold=0.1):

    mapSmooth = cv2.GaussianBlur(probMap, (3, 3), 0, 0)
    mapMask = np.uint8(mapSmooth>threshold)
    keypoints = []
    contours = None
    try:
        #OpenCV4.x
        contours, _ = cv2.findContours(mapMask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    except:
        #OpenCV3.x
        _, contours, _ = cv2.findContours(mapMask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

    for cnt in contours:
        blobMask = np.zeros(mapMask.shape)
        blobMask = cv2.fillConvexPoly(blobMask, cnt, 1)
        maskedProbMap = mapSmooth * blobMask
        _, maxVal, _, maxLoc = cv2.minMaxLoc(maskedProbMap)
        keypoints.append(maxLoc + (probMap[maxLoc[1], maxLoc[0]],))

    return keypoints


def getValidPairs(outputs, w, h):
    valid_pairs = []
    invalid_pairs = []
    n_interp_samples = 10
    paf_score_th = 0.1
    conf_th = 0.7

    for k in range(len(mapIdx)):
        pafA = outputs[0, mapIdx[k][0], :, :]
        pafB = outputs[0, mapIdx[k][1], :, :]
        pafA = cv2.resize(pafA, (w, h))
        pafB = cv2.resize(pafB, (w, h))

        candA = detected_keypoints[POSE_PAIRS[k][0]]
        candB = detected_keypoints[POSE_PAIRS[k][1]]
        nA = len(candA)
        nB = len(candB)

        if( nA != 0 and nB != 0):
            valid_pair = np.zeros((0,3))
            for i in range(nA):
                max_j=-1
                maxScore = -1
                found = 0
                for j in range(nB):
                    d_ij = np.subtract(candB[j][:2], candA[i][:2])
                    norm = np.linalg.norm(d_ij)
                    if norm:
                        d_ij = d_ij / norm
                    else:
                        continue
                    interp_coord = list(zip(np.linspace(candA[i][0], candB[j][0], num=n_interp_samples),
                                            np.linspace(candA[i][1], candB[j][1], num=n_interp_samples)))
                    paf_interp = []
                    for k in range(len(interp_coord)):
                        paf_interp.append([pafA[int(round(interp_coord[k][1])), int(round(interp_coord[k][0]))],
                                           pafB[int(round(interp_coord[k][1])), int(round(interp_coord[k][0]))] ])
                    paf_scores = np.dot(paf_interp, d_ij)
                    avg_paf_score = sum(paf_scores)/len(paf_scores)

                    if ( len(np.where(paf_scores > paf_score_th)[0]) / n_interp_samples ) > conf_th :
                        if avg_paf_score > maxScore:
                            max_j = j
                            maxScore = avg_paf_score
                            found = 1
                if found:
                    valid_pair = np.append(valid_pair, [[candA[i][3], candB[max_j][3], maxScore]], axis=0)

            valid_pairs.append(valid_pair)
        else:
            invalid_pairs.append(k)
            valid_pairs.append([])
    return valid_pairs, invalid_pairs


def getPersonwiseKeypoints(valid_pairs, invalid_pairs):
    personwiseKeypoints = -1 * np.ones((0, 19))

    for k in range(len(mapIdx)):
        if k not in invalid_pairs:
            partAs = valid_pairs[k][:,0]
            partBs = valid_pairs[k][:,1]
            indexA, indexB = np.array(POSE_PAIRS[k])

            for i in range(len(valid_pairs[k])):
                found = 0
                person_idx = -1
                for j in range(len(personwiseKeypoints)):
                    if personwiseKeypoints[j][indexA] == partAs[i]:
                        person_idx = j
                        found = 1
                        break

                if found:
                    personwiseKeypoints[person_idx][indexB] = partBs[i]
                    personwiseKeypoints[person_idx][-1] += keypoints_list[partBs[i].astype(int), 2] + valid_pairs[k][i][2]

                elif not found and k < 17:
                    row = -1 * np.ones(19)
                    row[indexA] = partAs[i]
                    row[indexB] = partBs[i]
                    row[-1] = sum(keypoints_list[valid_pairs[k][i,:2].astype(int), 2]) + valid_pairs[k][i][2]
                    personwiseKeypoints = np.vstack([personwiseKeypoints, row])
    return personwiseKeypoints


width  = 320
height = 240
fps = ""
framecount = 0
time1 = 0
elapsedTime = 0
index_void = 2
num_threads = 4

keypointsMapping = ['Nose', 'Neck', 'R-Sho', 'R-Elb', 'R-Wr', 'L-Sho', 'L-Elb', 'L-Wr', 'R-Hip', 'R-Knee', 'R-Ank', 'L-Hip', 'L-Knee', 'L-Ank', 'R-Eye', 'L-Eye', 'R-Ear', 'L-Ear']
POSE_PAIRS = [[1,2], [1,5], [2,3], [3,4], [5,6], [6,7], [1,8], [8,9], [9,10], [1,11], [11,12], [12,13], [1,0], [0,14], [14,16], [0,15], [15,17], [2,17], [5,16]]
mapIdx = [[31,32], [39,40], [33,34], [35,36], [41,42], [43,44], [19,20], [21,22], [23,24], [25,26], [27,28], [29,30], [47,48], [49,50], [53,54], [51,52], [55,56], [37,38], [45,46]]
colors = [[0,100,255], [0,100,255], [0,255,255], [0,100,255], [0,255,255], [0,100,255], [0,255,0], [255,200,100], [255,0,255], [0,255,0], [255,200,100], [255,0,255], [0,0,255], [255,0,0], [200,200,0], [255,0,0], [200,200,0], [0,0,0]]

cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FPS, 30)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)

#input_details  = [{'shape': array([ 1, 368, 432, 3], dtype=int32), 'quantization': (0.0078125, 128), 'index': 493, 'dtype': <class 'numpy.uint8'>, 'name': 'image'}]
#output_details = [{'shape': array([ 1, 46, 54, 57], dtype=int32), 'quantization': (0.0235294122248888, 0), 'index': 490, 'dtype': <class 'numpy.uint8'>, 'name': 'Openpose/concat_stage7'}]

try:
    # Tensorflow v1.13.0+

    #interpreter = tf.lite.Interpreter(model_path="models/train/test/tflite/mobilenet_v2_1.4_224/output_tflite_graph.tflite")
    interpreter = tf.lite.Interpreter(model_path="models/train/test/tflite/mobilenet_v2_0.5_224/output_tflite_graph.tflite")
except:
    # Tensorflow v1.12.0-

    #interpreter = tf.contrib.lite.Interpreter(model_path="models/train/test/tflite/mobilenet_v2_1.4_224/output_tflite_graph.tflite")
    interpreter = tf.contrib.lite.Interpreter(model_path="models/train/test/tflite/mobilenet_v2_0.5_224/output_tflite_graph.tflite")

interpreter.allocate_tensors()
interpreter.set_num_threads(int(num_threads))
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
input_shape = input_details[0]['shape']

h = input_details[0]['shape'][1] #368
w = input_details[0]['shape'][2] #432

threshold = 0.1
nPoints = 18

while True:
    t1 = time.perf_counter()
    ret, color_image = cap.read()
    if not ret:
        break

    # Resize image
    colw = color_image.shape[1]
    colh = color_image.shape[0]
    new_w = int(colw * min(w/colw, h/colh))
    new_h = int(colh * min(w/colw, h/colh))
    resized_image = cv2.resize(color_image, (new_w, new_h), interpolation = cv2.INTER_CUBIC)
    canvas = np.full((h, w, 3), 128)
    canvas[(h - new_h)//2:(h - new_h)//2 + new_h,(w - new_w)//2:(w - new_w)//2 + new_w, :] = resized_image
    prepimg = canvas
    prepimg = prepimg[np.newaxis, :, :, :] # Batch size axis add

    # Estimation
    interpreter.set_tensor(input_details[0]['index'], np.array(prepimg, dtype=np.uint8))
    interpreter.invoke()
    outputs = interpreter.get_tensor(output_details[0]['index']) #(1, 46, 54, 57)
    outputs = outputs.transpose((0, 3, 1, 2)) #(1, 57, 46, 54)

    # View
    detected_keypoints = []
    keypoints_list = np.zeros((0, 3))
    keypoint_id = 0

    for part in range(nPoints):
        probMap = outputs[0, part, :, :]
        probMap = cv2.resize(probMap, (canvas.shape[1], canvas.shape[0])) # (432, 368)
        keypoints = getKeypoints(probMap, threshold)
        keypoints_with_id = []

        for i in range(len(keypoints)):
            keypoints_with_id.append(keypoints[i] + (keypoint_id,))
            keypoints_list = np.vstack([keypoints_list, keypoints[i]])
            keypoint_id += 1

        detected_keypoints.append(keypoints_with_id)

    frameClone = np.uint8(canvas.copy())
    for i in range(nPoints):
        for j in range(len(detected_keypoints[i])):
            cv2.circle(frameClone, detected_keypoints[i][j][0:2], 5, colors[i], -1, cv2.LINE_AA)

    valid_pairs, invalid_pairs = getValidPairs(outputs, w, h)
    personwiseKeypoints = getPersonwiseKeypoints(valid_pairs, invalid_pairs)

    for i in range(17):
        for n in range(len(personwiseKeypoints)):
            index = personwiseKeypoints[n][np.array(POSE_PAIRS[i])]
            if -1 in index:
                continue
            B = np.int32(keypoints_list[index.astype(int), 0])
            A = np.int32(keypoints_list[index.astype(int), 1])
            cv2.line(frameClone, (B[0], A[0]), (B[1], A[1]), colors[i], 3, cv2.LINE_AA)

    cv2.putText(frameClone, fps, (w-170,15), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (38,0,255), 1, cv2.LINE_AA)

    cv2.namedWindow("USB Camera", cv2.WINDOW_AUTOSIZE)
    cv2.imshow("USB Camera" , frameClone)

    if cv2.waitKey(1)&0xFF == ord('q'):
        break

    # FPS calculation
    framecount += 1
    if framecount >= 5:
        fps = "(Playback) {:.1f} FPS".format(time1/15)
        framecount = 0
        time1 = 0
    t2 = time.perf_counter()
    elapsedTime = t2-t1
    time1 += 1/elapsedTime

cap.release()
cv2.destroyAllWindows()

４．Finally

GPD Pocket2 というモバイルPCを所有していますが、正直に言うと、こういう小型PCをライン側などにそっと置いておくだけで色々できそうな気がしています。 この小型PCは Core m3 を搭載しています。 デフォルトWindows10 ですが、 Ubuntu化が可能です。
当然ながら、突き抜けた高精度を求める場合は、GPU あるいは クラウドパワーで巨大なモデルをブンブンする必要があります。
http://gpdjapan.com/gpd_pocket_info/
201808281449_7.jpg
トレーニング用のスクリプト、Tensorflowのチェックポイント、推論用スクリプト、カスタマイズ済みTensorflow、の全てはGithubのリポジトリにコミット済みですので、 使えそうな機会があればご自由にご利用ください。
次は、Jetson Nano か TPU を使用した Pose Estimation にチャレンジしてみたいと思います。

最近、マシンパワーにものを言わせてGPUでブイブイするのがどれほど現実的な話なのか疑問に感じていたり。。。ボソッ :expressionless:

５．Appendix

７月に名古屋で ディープラーニングガジェット品評会 を開催いたします。 ご興味が有る方は こちらの記事 をご覧ください。
只今、発表者の事前予約を受付中です。

17
11
2

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
Sign upLogin
17
11

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?