AlphaPose姿态估计 只输出骨架 源码修改 相关记录

最新推荐文章于 2023-06-19 18:55:38 发布
最新推荐文章于 2023-06-19 18:55:38 发布
阅读量3.7k 收藏 64
点赞数 9
分类专栏： 笔记
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/souyan1991/article/details/112976260
版权
本文介绍了如何修改AlphaPose的视频输出形式，将连线颜色调整为白色，裁剪视频尺寸并降低关键点检测阈值，以增强人体姿态预测数据的质量。同时解决了无检测人时背景显示的问题，提升了处理速度和可视化效果。
摘要由CSDN通过智能技术生成

代码地址：https://github.com/MVIG-SJTU/AlphaPose

文章目录


先贴一下运行环境：
在这里插入图片描述
在这里插入图片描述
由于项目需要制作一系列人体姿态预测的预测数据，但是仅想要保留PoseFlow，去除原视频以提高处理速度。

1.修改视频输出形式

首先要按照输出的内容查找保存视频的代码。试着运行了一下demo，终端输出和视频结果如下：

100%|█████████████████████████████| 330/330 [00:42<00:00,  7.83it/s]
===========================> Finish Model Running.
===========================> Rendering remaining 0 images in the queue...
Results have been written to json.

在这里插入图片描述

根据第三行内容找到对应代码在demo_inference.py的287和273行。是通过print_finish_info输出的。查看进入条件，由于sp==False，所以确定位置为代码第273行的输出。

        print_finish_info()
        while(writer.running()):
            time.sleep(1)
            print('===========================> Rendering remaining ' + str(writer.count()) + ' images in the queue...')
        writer.stop()

往前查找207行附近发现是通过一个叫DataWriter的类执行了写入视频，这个类封装在了在alphapose/utils/writer.py

    if args.save_video and mode != 'image':
        from alphapose.utils.writer import DEFAULT_VIDEO_SAVE_OPT as video_save_opt
        if mode == 'video':
            video_save_opt['savepath'] = os.path.join(args.outputpath, 'AlphaPose_' + os.path.basename(input_source))
        else:
            video_save_opt['savepath'] = os.path.join(args.outputpath, 'AlphaPose_webcam' + str(input_source) + '.mp4')
        video_save_opt.update(det_loader.videoinfo)
        writer = DataWriter(cfg, args, save_video=True, video_save_opt=video_save_opt, queueSize=queueSize).start()
    else:
        writer = DataWriter(cfg, args, save_video=False, queueSize=queueSize).start()

writer.py中一路寻找opencv中的视频写入函数：VideoWriter().write()发现其配封装在了函数write_image()中，同时根据输入条件判定–save_video==True确定代码位置。141行附近：

                if self.opt.save_img or self.save_video or self.opt.vis:
                    if hm_data.size()[1] == 49:
                        from alphapose.utils.vis import vis_frame_dense as vis_frame
                    elif self.opt.vis_fast:
                        from alphapose.utils.vis import vis_frame_fast as vis_frame
                    else:
                        from alphapose.utils.vis import vis_frame
                    img = vis_frame(orig_img, result, self.opt)
                    self.write_image(img, im_name, stream=stream if self.save_video else None)

写入的视频帧img=vis_frame()，我们需要对这个作出修改。通过查看封装在alphapose/utils/vis.py中的vis_frame()函数发现，其中orig_img和result分别是原视频帧图像和预测图。代码如下：

def vis_frame(frame, im_res, opt, format='coco'):
    '''
    frame: frame image
    im_res: im_res of predictions
    format: coco or mpii

    return rendered image
    '''
    kp_num = 17
    if len(im_res['result']) > 0:
    	kp_num = len(im_res['result'][0]['keypoints'])

    if kp_num == 17:
        if format == 'coco':
            l_pair = [
                (0, 1), (0, 2), (1, 3), (2, 4),  # Head
                (5, 6), (5, 7), (7, 9), (6, 8), (8, 10),
                (17, 11), (17, 12),  # Body
                (11, 13), (12, 14), (13, 15), (14, 16)
            ]

            p_color = [(0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0),  # Nose, LEye, REye, LEar, REar
                       (77, 255, 255), (77, 255, 204), (77, 204, 255), (191, 255, 77), (77, 191, 255), (191, 255, 77),  # LShoulder, RShoulder, LElbow, RElbow, LWrist, RWrist
                       (204, 77, 255), (77, 255, 204), (191, 77, 255), (77, 255, 191), (127, 77, 255), (77, 255, 127), (0, 255, 255)]  # LHip, RHip, LKnee, Rknee, LAnkle, RAnkle, Neck
            line_color = [(0, 215, 255), (0, 255, 204), (0, 134, 255), (0, 255, 50),
                          (77, 255, 222), (77, 196, 255), (77, 135, 255), (191, 255, 77), (77, 255, 77),
                          (77, 222, 255), (255, 156, 127),
                          (0, 127, 255), (255, 127, 77), (0, 77, 255), (255, 77, 36)]
        elif format == 'mpii':
            l_pair = [
                (8, 9), (11, 12), (11, 10), (2, 1), (1, 0),
                (13, 14), (14, 15), (3, 4), (4, 5),
                (8, 7), (7, 6), (6, 2), (6, 3), (8, 12), (8, 13)
            ]
            p_color = [PURPLE, BLUE, BLUE, RED, RED, BLUE, BLUE, RED, RED, PURPLE, PURPLE, PURPLE, RED, RED, BLUE, BLUE]
            line_color = [PURPLE, BLUE, BLUE, RED, RED, BLUE, BLUE, RED, RED, PURPLE, PURPLE, RED, RED, BLUE, BLUE]
        else:
            raise NotImplementedError
    elif kp_num == 136:
        l_pair = [
            (0, 1), (0, 2), (1, 3), (2, 4),  # Head
            (5, 18), (6, 18), (5, 7), (7, 9), (6, 8), (8, 10),# Body
            (17, 18), (18, 19), (19, 11), (19, 12),
            (11, 13), (12, 14), (13, 15), (14, 16),
            (20, 24), (21, 25), (23, 25), (22, 24), (15, 24), (16, 25),# Foot
            (26, 27),(27, 28),(28, 29),(29, 30),(30, 31),(31, 32),(32, 33),(33, 34),(34, 35),(35, 36),(36, 37),(37, 38),#Face
            (38, 39),(39, 40),(40, 41),(41, 42),(43, 44),(44, 45),(45, 46),(46, 47),(48, 49),(49, 50),(50, 51),(51, 52),#Face
            (53, 54),(54, 55),(55, 56),(57, 58),(58, 59),(59, 60),(60, 61),(62, 63),(63, 64),(64, 65),(65, 66),(66, 67),#Face
            (68, 69),(69, 70),(70, 71),(71, 72),(72, 73),(74, 75),(75, 76),(76, 77),(77, 78),(78, 79),(79, 80),(80, 81),#Face
            (81, 82),(82, 83),(83, 84),(84, 85),(85, 86),(86, 87),(87, 88),(88, 89),(89, 90),(90, 91),(91, 92),(92, 93),#Face
            (94,95),(95,96),(96,97),(97,98),(94,99),(99,100),(100,101),(101,102),(94,103),(103,104),(104,105),#LeftHand
            (105,106),(94,107),(107,108),(108,109),(109,110),(94,111),(111,112),(112,113),(113,114),#LeftHand
            (115,116),(116,117),(117,118),(118,119),(115,120),(120,121),(121,122),(122,123),(115,124),(124,125),#RightHand
            (125,126),(126,127),(115,128),(128,129),(129,130),(130,131),(115,132),(132,133),(133,134),(134,135)#RightHand
        ]
        p_color = [(0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0),  # Nose, LEye, REye, LEar, REar
                   (77, 255, 255), (77, 255, 204), (77, 204, 255), (191, 255, 77), (77, 191, 255), (191, 255, 77),  # LShoulder, RShoulder, LElbow, RElbow, LWrist, RWrist
                   (204, 77, 255), (77, 255, 204), (191, 77, 255), (77, 255, 191), (127, 77, 255), (77, 255, 127),  # LHip, RHip, LKnee, Rknee, LAnkle, RAnkle, Neck
                   (77, 255, 255), (0, 255, 255), (77, 204, 255),  # head, neck, shoulder
                   (0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0), (77, 255, 255)] # foot
    
        line_color = [(0, 215, 255), (0, 255, 204), (0, 134, 255), (0, 255, 50),
                      (0, 255, 102), (77, 255, 222), (77, 196, 255), (77, 135, 255), (191, 255, 77), (77, 255, 77),
                      (77, 191, 255), (204, 77, 255), (77, 222, 255), (255, 156, 127),
                      (0, 127, 255), (255, 127, 77), (0, 77, 255), (255, 77, 36), 
                      (0, 77, 255), (0, 77, 255), (0, 77, 255), (0, 77, 255), (255, 156, 127), (255, 156, 127)]
    elif kp_num == 26:
        l_pair = [
            (0, 1), (0, 2), (1, 3), (2, 4),  # Head
            (5, 18), (6, 18), (5, 7), (7, 9), (6, 8), (8, 10),# Body
            (17, 18), (18, 19), (19, 11), (19, 12),
            (11, 13), (12, 14), (13, 15), (14, 16),
            (20, 24), (21, 25), (23, 25), (22, 24), (15, 24), (16, 25),# Foot
        ]
        p_color = [(0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0),  # Nose, LEye, REye, LEar, REar
                   (77, 255, 255), (77, 255, 204), (77, 204, 255), (191, 255, 77), (77, 191, 255), (191, 255, 77),  # LShoulder, RShoulder, LElbow, RElbow, LWrist, RWrist
                   (204, 77, 255), (77, 255, 204), (191, 77, 255), (77, 255, 191), (127, 77, 255), (77, 255, 127),  # LHip, RHip, LKnee, Rknee, LAnkle, RAnkle, Neck
                   (77, 255, 255), (0, 255, 255), (77, 204, 255),  # head, neck, shoulder
                   (0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0), (77, 255, 255)] # foot
    
        line_color = [(0, 215, 255), (0, 255, 204), (0, 134, 255), (0, 255, 50),
                      (0, 255, 102), (77, 255, 222), (77, 196, 255), (77, 135, 255), (191, 255, 77), (77, 255, 77),
                      (77, 191, 255), (204, 77, 255), (77, 222, 255), (255, 156, 127),
                      (0, 127, 255), (255, 127, 77), (0, 77, 255), (255, 77, 36), 
                      (0, 77, 255), (0, 77, 255), (0, 77, 255), (0, 77, 255), (255, 156, 127), (255, 156, 127)]
    else:
        raise NotImplementedError
    # im_name = os.path.basename(im_res['imgname'])
    img = frame.copy()
    height, width = img.shape[:2]
    for human in im_res['result']:
        part_line = {}
        kp_preds = human['keypoints']
        kp_scores = human['kp_score']
        if kp_num == 17:
            kp_preds = torch.cat((kp_preds, torch.unsqueeze((kp_preds[5, :] + kp_preds[6, :]) / 2, 0)))
            kp_scores = torch.cat((kp_scores, torch.unsqueeze((kp_scores[5, :] + kp_scores[6, :]) / 2, 0)))
        if opt.tracking:
            color = get_color_fast(int(abs(human['idx'])))
        else:
            color = BLUE

        # Draw bboxes
        if opt.showbox:
            if 'box' in human.keys():
                bbox = human['box']
                bbox = [bbox[0], bbox[0]+bbox[2], bbox[1], bbox[1]+bbox[3]]#xmin,xmax,ymin,ymax
            else:
                from trackers.PoseFlow.poseflow_infer import get_box
                keypoints = []
                for n in range(kp_scores.shape[0]):
                    keypoints.append(float(kp_preds[n, 0]))
                    keypoints.append(float(kp_preds[n, 1]))
                    keypoints.append(float(kp_scores[n]))
                bbox = get_box(keypoints, height, width)
            # color = get_color_fast(int(abs(human['idx'][0][0])))
            cv2.rectangle(img, (int(bbox[0]), int(bbox[2])), (int(bbox[1]),int(bbox[3])), color, 1)
            if opt.tracking:
                cv2.putText(img, str(human['idx']), (int(bbox[0]), int((bbox[2] + 26))), DEFAULT_FONT, 1, BLACK, 2)

        # Draw keypoints
        vis_thres = 0.05 if kp_num == 136 else 0.4
        for n in range(kp_scores.shape[0]):
            if kp_scores[n] <= vis_thres:
                continue
            cor_x, cor_y = int(kp_preds[n, 0]), int(kp_preds[n, 1])
            part_line[n] = (int(cor_x), int(cor_y))
            bg = img.copy()
            if n < len(p_color):
                if opt.tracking:
                    cv2.circle(bg, (int(cor_x), int(cor_y)), 2, color, -1)
                else:
                    cv2.circle(bg, (int(cor_x), int(cor_y)), 2, p_color[n], -1)
            else:
                cv2.circle(bg, (int(cor_x), int(cor_y)), 1, (255,255,255), 2)
            # Now create a mask of logo and create its inverse mask also
            if n < len(p_color):
                transparency = float(max(0, min(1, kp_scores[n])))
            else:
                transparency = float(max(0, min(1, kp_scores[n]*2)))
            img = cv2.addWeighted(bg, transparency, img, 1 - transparency, 0)
        # Draw limbs
        for i, (start_p, end_p) in enumerate(l_pair):
            if start_p in part_line and end_p in part_line:
                start_xy = part_line[start_p]
                end_xy = part_line[end_p]
                bg = img.copy()

                X = (start_xy[0], end_xy[0])
                Y = (start_xy[1], end_xy[1])
                mX = np.mean(X)
                mY = np.mean(Y)
                length = ((Y[0] - Y[1]) ** 2 + (X[0] - X[1]) ** 2) ** 0.5
                angle = math.degrees(math.atan2(Y[0] - Y[1], X[0] - X[1]))
                stickwidth = (kp_scores[start_p] + kp_scores[end_p]) + 1
                polygon = cv2.ellipse2Poly((int(mX), int(mY)), (int(length/2), int(stickwidth)), int(angle), 0, 360, 1)
                if i < len(line_color):
                    if opt.tracking:
                        cv2.fillConvexPoly(bg, polygon, color)
                    else:
                        cv2.fillConvexPoly(bg, polygon, line_color[i])
                else:
                    cv2.line(bg, start_xy, end_xy, (255,255,255), 1)
                if n < len(p_color):
                    transparency = float(max(0, min(1, 0.5 * (kp_scores[start_p] + kp_scores[end_p])-0.1)))
                else:
                    transparency = float(max(0, min(1, (kp_scores[start_p] + kp_scores[end_p]))))

                #transparency = float(max(0, min(1, 0.5 * (kp_scores[start_p] + kp_scores[end_p])-0.1)))
                img = cv2.addWeighted(bg, transparency, img, 1 - transparency, 0)
    return img

输出的img在120行和269行：

img=frame.copy()

如果想把输出改成单纯的poseflow可以用以下代码替换。至此实现输出的修改。最后运行看下结果：

    height, width = frame.shape[:2]
    img=np.zeros((height, width, 3), dtype=np.uint8) #背景改成黑色画布

在这里插入图片描述

修改骨骼连线属性

1.修改骨骼连线颜色

骨骼连线和关键点的各种参数都写在了./alphapose/utils/vis.py的vis_frame_fast函数和vis_frame函数下面。
这个文件是关于处理可视化结果的操作。这里贴上修改完的代码：

import math
import time

import cv2
import numpy as np
import torch

RED = (0, 0, 255)
GREEN = (0, 255, 0)
BLUE = (255, 0, 0)
CYAN = (255, 255, 0)
YELLOW = (0, 255, 255)
ORANGE = (0, 165, 255)
PURPLE = (255, 0, 255)
WHITE = (255, 255, 255)
BLACK = (0, 0, 0)

DEFAULT_FONT = cv2.FONT_HERSHEY_SIMPLEX


def get_color(idx):
    idx = idx * 3
    color = ((37 * idx) % 255, (17 * idx) % 255, (29 * idx) % 255)

    return color


def get_color_fast(idx):
    color_pool = [RED, GREEN, BLUE, CYAN, YELLOW, ORANGE, PURPLE, WHITE]
    color = color_pool[idx % 8]

    return color


#上面vis_frame_fast的内容和下面的vis_frame类似，是快速进行可视化的操作。
def vis_frame(frame, im_res, opt, format='coco'):
    '''
    frame: frame image
    im_res: im_res of predictions
    format: coco or mpii

    return rendered image
    '''
    kp_num = 17
    if len(im_res['result']) > 0:
    	kp_num = len(im_res['result'][0]['keypoints'])

    #kp_num = 17 将关键点固定为17个

    if kp_num == 17:
        if format == 'coco':
            l_pair = [
                (0, 1), (0, 2), (1, 3), (2, 4),  # Head
                (5, 6), (5, 7), (7, 9), (6, 8), (8, 10),
                (17, 11), (17, 12),  # Body
                (11, 13), (12, 14), (13, 15), (14, 16)
            ]
			#坐标点和连线的默认颜色，我这里修改成白色了
            # p_color = [(0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0),  # Nose, LEye, REye, LEar, REar
            #            (77, 255, 255), (77, 255, 204), (77, 204, 255), (191, 255, 77), (77, 191, 255), (191, 255, 77),  # LShoulder, RShoulder, LElbow, RElbow, LWrist, RWrist
            #            (204, 77, 255), (77, 255, 204), (191, 77, 255), (77, 255, 191), (127, 77, 255), (77, 255, 127), (0, 255, 255)] #18  # LHip, RHip, LKnee, Rknee, LAnkle, RAnkle, Neck
            # line_color = [(0, 215, 255), (0, 255, 204), (0, 134, 255), (0, 255, 50),
            #               (77, 255, 222), (77, 196, 255), (77, 135, 255), (191, 255, 77), (77, 255, 77),
            #               (77, 222, 255), (255, 156, 127),
            #               (0, 127, 255), (255, 127, 77), (0, 77, 255), (255, 77, 36)] #15

            p_color = [(255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255),
                       (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255),
                       (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255),
                       (255, 255, 255), (255, 255, 255), (255, 255, 255), ] #18
            line_color = [(255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255),
                          (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255),
                          (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255), (255, 255, 255)]


        elif format == 'mpii':
            l_pair = [
                (8, 9), (11, 12), (11, 10), (2, 1), (1, 0),
                (13, 14), (14, 15), (3, 4), (4, 5),
                (8, 7), (7, 6), (6, 2), (6, 3), (8, 12), (8, 13)
            ]
            p_color = [PURPLE, BLUE, BLUE, RED, RED, BLUE, BLUE, RED, RED, PURPLE, PURPLE, PURPLE, RED, RED, BLUE, BLUE]
            line_color = [PURPLE, BLUE, BLUE, RED, RED, BLUE, BLUE, RED, RED, PURPLE, PURPLE, RED, RED, BLUE, BLUE]
        else:
            raise NotImplementedError
    elif kp_num == 136:
        l_pair = [
            (0, 1), (0, 2), (1, 3), (2, 4),  # Head
            (5, 18), (6, 18), (5, 7), (7, 9), (6, 8), (8, 10),# Body
            (17, 18), (18, 19), (19, 11), (19, 12),
            (11, 13), (12, 14), (13, 15), (14, 16),
            (20, 24), (21, 25), (23, 25), (22, 24), (15, 24), (16, 25),# Foot
            (26, 27),(27, 28),(28, 29),(29, 30),(30, 31),(31, 32),(32, 33),(33, 34),(34, 35),(35, 36),(36, 37),(37, 38),#Face
            (38, 39),(39, 40),(40, 41),(41, 42),(43, 44),(44, 45),(45, 46),(46, 47),(48, 49),(49, 50),(50, 51),(51, 52),#Face
            (53, 54),(54, 55),(55, 56),(57, 58),(58, 59),(59, 60),(60, 61),(62, 63),(63, 64),(64, 65),(65, 66),(66, 67),#Face
            (68, 69),(69, 70),(70, 71),(71, 72),(72, 73),(74, 75),(75, 76),(76, 77),(77, 78),(78, 79),(79, 80),(80, 81),#Face
            (81, 82),(82, 83),(83, 84),(84, 85),(85, 86),(86, 87),(87, 88),(88, 89),(89, 90),(90, 91),(91, 92),(92, 93),#Face
            (94,95),(95,96),(96,97),(97,98),(94,99),(99,100),(100,101),(101,102),(94,103),(103,104),(104,105),#LeftHand
            (105,106),(94,107),(107,108),(108,109),(109,110),(94,111),(111,112),(112,113),(113,114),#LeftHand
            (115,116),(116,117),(117,118),(118,119),(115,120),(120,121),(121,122),(122,123),(115,124),(124,125),#RightHand
            (125,126),(126,127),(115,128),(128,129),(129,130),(130,131),(115,132),(132,133),(133,134),(134,135)#RightHand
        ]
        p_color = [(0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0),  # Nose, LEye, REye, LEar, REar
                   (77, 255, 255), (77, 255, 204), (77, 204, 255), (191, 255, 77), (77, 191, 255), (191, 255, 77),  # LShoulder, RShoulder, LElbow, RElbow, LWrist, RWrist
                   (204, 77, 255), (77, 255, 204), (191, 77, 255), (77, 255, 191), (127, 77, 255), (77, 255, 127),  # LHip, RHip, LKnee, Rknee, LAnkle, RAnkle, Neck
                   (77, 255, 255), (0, 255, 255), (77, 204, 255),  # head, neck, shoulder
                   (0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0), (77, 255, 255)] # foot
    
        line_color = [(0, 215, 255), (0, 255, 204), (0, 134, 255), (0, 255, 50),
                      (0, 255, 102), (77, 255, 222), (77, 196, 255), (77, 135, 255), (191, 255, 77), (77, 255, 77),
                      (77, 191, 255), (204, 77, 255), (77, 222, 255), (255, 156, 127),
                      (0, 127, 255), (255, 127, 77), (0, 77, 255), (255, 77, 36), 
                      (0, 77, 255), (0, 77, 255), (0, 77, 255), (0, 77, 255), (255, 156, 127), (255, 156, 127)]
    elif kp_num == 26:
        l_pair = [
            (0, 1), (0, 2), (1, 3), (2, 4),  # Head
            (5, 18), (6, 18), (5, 7), (7, 9), (6, 8), (8, 10),# Body
            (17, 18), (18, 19), (19, 11), (19, 12),
            (11, 13), (12, 14), (13, 15), (14, 16),
            (20, 24), (21, 25), (23, 25), (22, 24), (15, 24), (16, 25),# Foot
        ]
        p_color = [(0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0),  # Nose, LEye, REye, LEar, REar
                   (77, 255, 255), (77, 255, 204), (77, 204, 255), (191, 255, 77), (77, 191, 255), (191, 255, 77),  # LShoulder, RShoulder, LElbow, RElbow, LWrist, RWrist
                   (204, 77, 255), (77, 255, 204), (191, 77, 255), (77, 255, 191), (127, 77, 255), (77, 255, 127),  # LHip, RHip, LKnee, Rknee, LAnkle, RAnkle, Neck
                   (77, 255, 255), (0, 255, 255), (77, 204, 255),  # head, neck, shoulder
                   (0, 255, 255), (0, 191, 255), (0, 255, 102), (0, 77, 255), (0, 255, 0), (77, 255, 255)] # foot
    
        line_color = [(0, 215, 255), (0, 255, 204), (0, 134, 255), (0, 255, 50),
                      (0, 255, 102), (77, 255, 222), (77, 196, 255), (77, 135, 255), (191, 255, 77), (77, 255, 77),
                      (77, 191, 255), (204, 77, 255), (77, 222, 255), (255, 156, 127),
                      (0, 127, 255), (255, 127, 77), (0, 77, 255), (255, 77, 36), 
                      (0, 77, 255), (0, 77, 255), (0, 77, 255), (0, 77, 255), (255, 156, 127), (255, 156, 127)] #24
    else:
        raise NotImplementedError

    #获取原视频的size
    height, width = frame.shape[:2]
    img = np.zeros((height, width, 3), dtype=np.uint8) #背景改成黑色画布
    #对检测结果进行可视化，获取每个点的置信度和预测值
    for human in im_res['result']:
        part_line = {}
        kp_preds = human['keypoints']
        kp_scores = human['kp_score']
        if kp_num == 17:
            kp_preds = torch.cat((kp_preds, torch.unsqueeze((kp_preds[5, :] + kp_preds[6, :]) / 2, 0)))
            kp_scores = torch.cat((kp_scores, torch.unsqueeze((kp_scores[5, :] + kp_scores[6, :]) / 2, 0)))
        if opt.tracking:
            color = get_color_fast(int(abs(human['idx'])))
        else:
            color = WHITE

        # Draw bboxes
        if opt.showbox:
            if 'box' in human.keys():
                bbox = human['box']
                bbox = [bbox[0], bbox[0]+bbox[2], bbox[1], bbox[1]+bbox[3]]#xmin,xmax,ymin,ymax
            else:
                from trackers.PoseFlow.poseflow_infer import get_box
                keypoints = []
                for n in range(kp_scores.shape[0]):
                    keypoints.append(float(kp_preds[n, 0]))
                    keypoints.append(float(kp_preds[n, 1]))
                    keypoints.append(float(kp_scores[n]))
                bbox = get_box(keypoints, height, width) ##视频帧高宽
            # color = get_color_fast(int(abs(human['idx'][0][0])))
            #bbox[x1,x2,y1,y2]
            cv2.rectangle(img, (int(bbox[0]), int(bbox[2])), (int(bbox[1]),int(bbox[3])), color, 1)
            if opt.tracking:
                cv2.putText(img, str(human['idx']), (int(bbox[0]), int((bbox[2] + 26))), DEFAULT_FONT, 1, BLACK, 2)

        # Draw keypoints
        vis_thres = 0.05 if kp_num == 136 else 0.35
        for n in range(kp_scores.shape[0]):
            if kp_scores[n] <= vis_thres:
                continue
            cor_x, cor_y = int(kp_preds[n, 0]), int(kp_preds[n, 1])
            part_line[n] = (int(cor_x), int(cor_y))
            bg = img.copy()
            if n < len(p_color):
                if opt.tracking:
                    cv2.circle(bg, (int(cor_x), int(cor_y)), 2, color, -1)
                else:
                    cv2.circle(bg, (int(cor_x), int(cor_y)), 2, p_color[n], -1)
            else:
                cv2.circle(bg, (int(cor_x), int(cor_y)), 1, (255,255,255), 2)
            # Now create a mask of logo and create its inverse mask also
            if n < len(p_color):
                transparency = float(max(0, min(1, kp_scores[n])))
            else:
                transparency = float(max(0, min(1, kp_scores[n]*2)))
            img = cv2.addWeighted(bg, transparency, img, 1 - transparency, 0)
        # Draw limbs
        for i, (start_p, end_p) in enumerate(l_pair):
            if start_p in part_line and end_p in part_line:
                start_xy = part_line[start_p]
                end_xy = part_line[end_p]
                bg = img.copy()

                X = (start_xy[0], end_xy[0])
                Y = (start_xy[1], end_xy[1])
                mX = np.mean(X)
                mY = np.mean(Y)
                length = ((Y[0] - Y[1]) ** 2 + (X[0] - X[1]) ** 2) ** 0.5
                angle = math.degrees(math.atan2(Y[0] - Y[1], X[0] - X[1]))
                stickwidth = (kp_scores[start_p] + kp_scores[end_p]) + 1
                polygon = cv2.ellipse2Poly((int(mX), int(mY)), (int(length/2), int(stickwidth)), int(angle), 0, 360, 1)
                if i < len(line_color):
                    if opt.tracking:
                        cv2.fillConvexPoly(bg, polygon, color)
                    else:
                        cv2.fillConvexPoly(bg, polygon, line_color[i])
                else:
                    cv2.line(bg, start_xy, end_xy, (255,255,255), 5) #主要修改的点，加粗线条到5，颜色改为白色
                if n < len(p_color):
                    transparency = float(max(0, min(1, 0.5 * (kp_scores[start_p] + kp_scores[end_p])-0.1)))
                else:
                    transparency = float(max(0, min(1, (kp_scores[start_p] + kp_scores[end_p]))))

                #transparency = float(max(0, min(1, 0.5 * (kp_scores[start_p] + kp_scores[end_p])-0.1)))
                img = cv2.addWeighted(bg, transparency, img, 1 - transparency, 0) #根据检测结果的置信度修改连线的权重（颜色的深浅）
    return img


def getTime(time1=0):
    if not time1:
        return time.time()
    else:
        interval = time.time() - time1
        return time.time(), interval

2.裁剪视频

为了方便训练，将视频裁剪成640*640大小，并取每段视频的时间轴上的中间部分（因为大部分数据的中间时刻都是行人经过画面中心的时刻），自己写了一段代码如下：

import os
import ffmpeg

input_dir = 'your path'
save_path = 'your path'

#裁剪成640*640大小
def crop(stream):
    stream = ffmpeg.crop(stream, x=640, y=320, width=640, height=640)

    return stream

#截取指定时间段的视频
def trim(stream, duration, length=2.0):
    start = float(duration) // 2
    end = start + length
    stream = ffmpeg.trim(stream, start=start, end=end)

    return stream

#对文件夹下的视频进行操作
video_list = os.listdir(input_dir)
i = 0
for file in video_list:
    i += 1
    file = input_dir + file
    duration = ffmpeg.probe(file)['format'].get('duration')

    stream = ffmpeg.input(file)
    stream = crop(stream)
    stream = trim(stream, duration)
    stream.output(save_path + 'train_%0.4d.mp4' % i).run()

修改后的视频效果如下：
thrd=0.4

3.修改关键点阈值

上面的结果有的时候会消失，原版的阈值为0.4，如果检测的score低于这个值就不输出point。根据自己的数据修改这个值，值太小连线就会很乱，因为confidence很低的结果也被画出来了。值太大gt也不会被画出来，导致“缺胳膊少腿”。还是需要根据自己的数据修改这个阈值。
代码在vis.py的函数vis_frame中：

# Draw keypoints
        vis_thres = 0.05 if kp_num == 136 else 0.35 <--这个值我修改为0.01
        for n in range(kp_scores.shape[0]):
            if kp_scores[n] <= vis_thres:
                continue
            cor_x, cor_y = int(kp_preds[n, 0]), int(kp_preds[n, 1])
            part_line[n] = (int(cor_x), int(cor_y))
            bg = img.copy()
            '''以下略'''

看下改成0.01后的结果：
thrd=0.01

4. 解决没有检测到人的时候背景不是黑色的问题

输出一些视频后发现，如果alpha pose在当前帧下没有检测到人的结果时，可视化时就会输出视频原帧，这种做法在输出【原帧】+【骨架】的画面时是没有问题的。现在我们想在没有检测到行人的时候输出纯黑色的背景，即【黑色背景】+【骨架】，就需要修改如下的代码：

'''代码位置在./alphapose/utils/writer.py的96行左右'''
# image channel RGB->BGR
            orig_img = np.array(orig_img, dtype=np.uint8)[:, :, ::-1]
            if boxes is None or len(boxes) == 0: #没有检测到行人预测框
                if self.opt.save_img or self.save_video or self.opt.vis:
                    height, width = orig_img.shape[:2] #获取原图像的size
                    black_img = np.zeros((height, width, 3), dtype=np.uint8) #背景改成黑色画布
                    self.write_image(black_img, im_name, stream=stream if self.save_video else None)

问题解决。

接下来如果还发现什么问题还会继续更新，欢迎讨论。
最后更新日期202104016

Abalizzw
关注
专栏目录
AlphaPose多路摄像头同步运行的修改方法
u013837919的博客
05-07 2769
作者只提供了单路摄像头的解决方法，我要实现多路摄像头，就做了很多尝试： 1.把多路信号放到一个列表，然后循环取每一路信号，放到一个batch里，送给模型识别,当然，这里没有封装，你也可以把地址放到opt.py里 url_dir = ['1', '2',‘3’,‘4’,'5','6'] lushu = len(url_dir) for n,url in enumera...
AlphaPose人体姿态识别之json文件解读
guomoyu的博客
08-16 4440
@[TOC](alphapose之json文件解读) alphapose之json文件解读 本文记录下res的json文件坐标系以及一些参数的意思，有不足之处望指正。 待解读例子： 例如有如下json文件 [{"image_id": "1.jpg", "category_id": 1, "keypoints": [654.6572265625, 88.31778717041016, 0.9614965915679932, 659.4932861328125, 78.64559173583984,0.93
python实现Alphapose骨骼关键点信息的提取
03-24
利用python实现Alphapose骨骼关键点信息的提取，将jison中的关键点信息存储到txt文件中。
alphapose输出2d骨架数据转化为每一个id在每一帧的2d骨架数据及检测框位置
qq_42316533的博客
07-18 790
import json import os import argparse from shutil import copy def is_300(data, points): # 判断当前id的总帧数是否为300帧 if len(data) > 300: for i in range(300, len(data)): del data[str(i)] if len(data) < 300: for i in range(
昇腾200、300、500、IVS1800上部署AlphaPose
qq_41384531的博客
08-25 532
这里写自定义目录标题欢迎使用Markdown编辑器新的改变功能快捷键合理的创建标题，有助于目录的生成如何改变文本的样式插入链接与图片如何插入一段漂亮的代码片生成一个适合你的列表创建一个表格设定内容居中、居左、居右SmartyPants创建一个自定义列表如何创建一个注脚注释也是必不可少的KaTeX数学公式新的甘特图功能，丰富你的文章UML 图表FLowchart流程图导出与导入导出导入 欢迎使用Markdown编辑器 你好！ 这是你第一次使用 Markdown编辑器 所展示的欢迎页。如果你想学习如何使用Mar
pose-实战人体姿态识别之AlphaPose-整理了一键执行的项目代码形式
最新发布
11-09
AlphaPose不仅能够检测出人体的关键点，还能构建出人体骨架图，从而实现对各种姿态的精确识别。这个技术在多个领域有着广泛的应用，包括智能监控、人机交互、运动分析等。 在智能监控中，人体姿态识别有助于自动...
AlphaPose：领先姿态估计系统的开源里程碑
AlphaPose系统专门设计用于多人姿态估计任务，能够在图像或视频中识别出多个人的骨架关键点。姿态估计是计算机视觉中的一个关键问题，对于理解场景中人的行为和动作具有重要作用。 2. COCO数据集与mAP（mean ...
AlphaPose-master.zip
09-02
AlphaPose是一款先进的开源人体姿态估计框架，其在准确性和效率方面优于OpenPose和Mask R-CNN等其他流行的人体姿态估计技术。这个压缩包“AlphaPose-master.zip”包含的源代码和相关资源使得用户能够利用深度学习...
AlphaPose轻量化SPPE训练代码
08-03
AlphaPose的轻量化单人姿态估计网络（SPPE）训练代码 使用步骤 1. 下载数据集放置到train_demo/data文件夹 2. pip install -r requirements.txt 3. 安装pycocotools和alphapose 3.1 cd train_demo 3.2 python setup.py install 3.3 python setup1.py build develop 4.开始训练 python train_demo/train.py 相关博客：https://blog.csdn.net/xzq1207105685/article/details/119331007
【AlphaPose】AlphaPose运行环境搭建与测试
weixin_42265958的博客
05-20 1万+
AlphaPose项目地址 AlphaPose是一种精确的多人姿势估计器，其运行环境的搭建相较于OpenPose而言更加简便。 搭建步骤 官方安装步骤 这里我选用官方推荐的conda进行安装测试。 安装anaconda 利用conda创建虚拟环境：更多创建虚拟环境的方法参考文章：Ubuntu下创建虚拟环境步骤 # 创建虚拟环境 conda create -n alphapose python=3.6 -y # 进入虚拟环境 source activate alphapose 安装pytorch
Alphapose算法
m0_55674689的博客
02-25 4562
alphapose算法理解
yolov8pose pt转onnx，并实现图片和视频推理
LKjoey的博客
06-19 10000
v8官网v8pose权重看自己设备配置和推理速度下载需要的模型。
OpenPose：实时多人2D姿态估计
黎国溥
05-29 1万+
原文地址：https://zhuanlan.zhihu.com/p/37526892 前言 前段时间很火的感人动画短片《Changing Batteries》讲了这样一个故事：独居的老奶奶收到儿子寄来的一个机器人，虽然鲜有语言的沟通，但是小机器人善于察言观色，不仅能在老奶奶口渴时为她端水、在老奶奶扫地时接过老奶奶的扫把，做力所能及的家务活，如果老奶奶在椅子上看电视睡着了，机器人还为她轻轻...
AlphaPose（RMPE）区域多人姿态估计理解
weixin_45060598的博客
04-24 5468
1.介绍 AlphaPose 是一个精准的多人姿态估计系统，是首个在 COCO 数据集上可达到 70+ mAP（72.3 mAP，高于 Mask-RCNN 8.2 个百分点），在 MPII 数据集上可达到 80+ mAP（82.1 mAP）的开源系统。为了能将同一个人的所有姿态关联起来，AlphaPose 还提供了一个称为 Pose Flow 的在线姿态跟踪器，这也是首个在 PoseTrack 挑战数据集上达到 60+ mAP（66.5 mAP）和 50+ MOTA（58.3 MOTA）的开源在线姿态跟踪
Alphapose论文代码详解
热门推荐
m0_45850873的博客
04-03 1万+
第1节 人体姿态估计的基本概念 第2节：Alphapose 2.1Alphapose简介 Alphapose是由上海交大大学的卢策吾团队和腾讯优图提出的一种区域多人姿态估计框架，论文全称：RMPE: Regional Multi-Person Pose Estimation，于2018年2月4日发表于ICCV，在MPII数据集上达到76.7mAP。 Alphapose是多人姿态估计方法，比单人姿态估计难度大，更具有挑战性。目前多人姿态估计的方法分为两类：一种是自顶向下，另一种是自底向上。自顶向下检测方法是先
多人姿态识别框架——AlphaPose
Wenrui Xie的博客
11-24 6288
多人姿态识别简介 人体姿态估计有两个主流方案：Two-step framework 和Part-based framework。第一种方案是检测环境中的每一个人体检测框，然后独立地去检测每一个人体区域的姿态（自顶向下的方法）。第二种方案是首先检测出环境中的所有肢体节点，然后进行拼接得到多人的骨架（自底向上的方法）。第一种方案，姿态检测准确度高度以来目标区域框检测的质量。第二种方案，如果两人离得十分近，容易出现模棱两可的情况，而且由于是依赖两个部件之间的关系，所以失去了对全局的信息获取。 AlphaPos
AlphaPose源代码学习之getPrediction()函数
Ji_HON的博客
12-10 1972
该函数的作用是从预测的热力图中获取关键点，位于eval.py中，完整代码如下 def getPrediction(hms, pt1, pt2, inpH, inpW, resH, resW): ''' 从热力图中获得关键点 ''' assert hms.dim() == 4, 'Score maps should be 4-dim' maxval, idx = torch.max(hms.view(hms.size(0), hms.size(1), -1), 2)
使用TensorRT对AlphaPose模型进行加速
oreo的博客
05-07 2750
最近刚完成使用TensorRT对AlphaPose人体姿态估计网络的加速处理，在这里记录一下大概的流程，具体代码我放在这里了。 目前主要有三种方式构建TensorRT的engine模型。 (1) 第一种是使用模型框架自带的方法生成engine模型，比如TensorFlow和MXNet框架支持直接转成TensorRT的engine模型，这种方式虽然便捷，但是运行效率较低； (2) 第二种是使用C++或者python的API直接构建检测模型，这种方式虽然效率上限高，但是实现步骤较为繁琐，兼容性较低，一旦原
Alphapose配置
NcepuKZH的博客
04-24 5738
对于人体骨架数据提取，从Openpose到Alphapose，对人少的场景都做出了很好的性能，港中文AAA12018的ST-GCN（https://arxiv.org/abs/1801.07455） 在Openpose的基础上，使用时空图卷积网络进行了简单动作识别，准确率有待提高。在Kinetics （7.5G）and NTU-RGBD（5.8G）数据上进行了相关测试。最近遇到了行人匹配的问题，由...
评论 10
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值