前言

先来看下实现效果：

上图展示了用yolov5作为检测器，DeepSort为追踪器实现了对车流量的统计并绘制了每辆车的运行轨迹。

一、整体目录结构

下图展示了项目的整体目录结构：

其中：

• deep_sort文件下为目标跟踪相关代码；
• weights文件夹下存放yolov5检测模型；
• demo.py针对读取的视频进行目标追踪
• objdetector.py封装的一个目标检测器，对视频中的物体进行检测
• objtracker.py封装了一个目标追踪器，对检测的物体进行追踪

二、DeepSort代码参数解释

deep_sort/configs/deep_sort.yaml文件里保存了DeepSort算法的配置参数：

这些参数依次的含义为：

• REID_CKPT:特征提取权重的目录路径
• MAX_DIST: 最大余弦距离，用于级联匹配，如果大于该阈值，则忽略
• MIN_CONFIDENCE:检测结果置信度阈值
• NMS_MAX_OVERLAP: 非极大抑制阈值，设置为1代表不进行抑制
• MAX_IOU_DISTANCE: 最大IOU阈值
• MAX_AGE:最大寿命，也就是经过MAX_AGE帧没有追踪到该物体，就将该轨迹变为删除态
• N_INIT: 最高击中次数，如果击中该次数，就由不确定态转为确定态
• NN_BUDGET: 最大保存特征帧数，如果超过该帧数，将进行滚动保存
  三、代码展示
  下面给出demo.py的代码：

import numpy as npimport objtracker
from objdetector import Detector
import cv2VIDEO_PATH = './video/test_traffic.mp4'if __name__ == '__main__':# 根据视频尺寸，填充供撞线计算使用的polygonwidth = 1920height = 1080mask_image_temp = np.zeros((height, width), dtype=np.uint8)# 用于记录轨迹信息pts = {}# 填充第一个撞线polygon（蓝色）list_pts_blue = [[204, 305], [227, 431], [605, 522], [1101, 464], [1900, 601], [1902, 495], [1125, 379], [604, 437],[299, 375], [267, 289]]ndarray_pts_blue = np.array(list_pts_blue, np.int32)polygon_blue_value_1 = cv2.fillPoly(mask_image_temp, [ndarray_pts_blue], color=1)polygon_blue_value_1 = polygon_blue_value_1[:, :, np.newaxis]# 填充第二个撞线polygon（黄色）mask_image_temp = np.zeros((height, width), dtype=np.uint8)list_pts_yellow = [[181, 305], [207, 442], [603, 544], [1107, 485], [1898, 625], [1893, 701], [1101, 568],[594, 637], [118, 483], [109, 303]]ndarray_pts_yellow = np.array(list_pts_yellow, np.int32)polygon_yellow_value_2 = cv2.fillPoly(mask_image_temp, [ndarray_pts_yellow], color=2)polygon_yellow_value_2 = polygon_yellow_value_2[:, :, np.newaxis]# 撞线检测用的mask，包含2个polygon，（值范围 0、1、2），供撞线计算使用polygon_mask_blue_and_yellow = polygon_blue_value_1 + polygon_yellow_value_2# 缩小尺寸，1920x1080->960x540polygon_mask_blue_and_yellow = cv2.resize(polygon_mask_blue_and_yellow, (width // 2, height // 2))# 蓝 色盘 b,g,rblue_color_plate = [255, 0, 0]# 蓝 polygon图片blue_image = np.array(polygon_blue_value_1 * blue_color_plate, np.uint8)# 黄 色盘yellow_color_plate = [0, 255, 255]# 黄 polygon图片yellow_image = np.array(polygon_yellow_value_2 * yellow_color_plate, np.uint8)# 彩色图片（值范围 0-255）color_polygons_image = blue_image + yellow_image# 缩小尺寸，1920x1080->960x540color_polygons_image = cv2.resize(color_polygons_image, (width // 2, height // 2))# list 与蓝色polygon重叠list_overlapping_blue_polygon = []# list 与黄色polygon重叠list_overlapping_yellow_polygon = []# 下行数量down_count = 0# 上行数量up_count = 0font_draw_number = cv2.FONT_HERSHEY_SIMPLEXdraw_text_postion = (int((width / 2) * 0.01), int((height / 2) * 0.05))# 实例化yolov5检测器detector = Detector()# 打开视频capture = cv2.VideoCapture(VIDEO_PATH)while True:# 读取每帧图片_, im = capture.read()if im is None:break# 缩小尺寸，1920x1080->960x540im = cv2.resize(im, (width // 2, height // 2))list_bboxs = []# 更新跟踪器output_image_frame, list_bboxs = objtracker.update(detector, im)# 输出图片output_image_frame = cv2.add(output_image_frame, color_polygons_image)if len(list_bboxs) > 0:# ----------------------判断撞线----------------------for item_bbox in list_bboxs:x1, y1, x2, y2, _, track_id = item_bbox# 撞线检测点，(x1，y1)，y方向偏移比例 0.0~1.0y1_offset = int(y1 + ((y2 - y1) * 0.5))x1_offset = int(x1 + ((x2 - x1) * 0.5))# 撞线的点y = y1_offsetx = x1_offset# 然后每检测出一个预测框，就将中心点加入队列center = (x, y)if track_id in pts:pts[track_id].append(center)else:pts[track_id] = []pts[track_id].append(center)thickness = 2cv2.circle(output_image_frame, (center), 1, [255, 255, 255], thickness)for j in range(1, len(pts[track_id])):if pts[track_id][j - 1] is None or pts[track_id][j] is None:continuecv2.line(output_image_frame, (pts[track_id][j - 1]), (pts[track_id][j]), [255, 255, 255], thickness)if polygon_mask_blue_and_yellow[y, x] == 1:# 如果撞 蓝polygonif track_id not in list_overlapping_blue_polygon:list_overlapping_blue_polygon.append(track_id)# 判断 黄polygon list里是否有此 track_id# 有此track_id，则认为是 UP (上行)方向if track_id in list_overlapping_yellow_polygon:# 上行+1up_count += 1print('up count:', up_count, ', up id:', list_overlapping_yellow_polygon)# 删除 黄polygon list 中的此idlist_overlapping_yellow_polygon.remove(track_id)elif polygon_mask_blue_and_yellow[y, x] == 2:# 如果撞 黄polygonif track_id not in list_overlapping_yellow_polygon:list_overlapping_yellow_polygon.append(track_id)# 判断 蓝polygon list 里是否有此 track_id# 有此 track_id，则 认为是 DOWN（下行）方向if track_id in list_overlapping_blue_polygon:# 下行+1down_count += 1print('down count:', down_count, ', down id:', list_overlapping_blue_polygon)# 删除 蓝polygon list 中的此idlist_overlapping_blue_polygon.remove(track_id)# ----------------------清除无用id----------------------list_overlapping_all = list_overlapping_yellow_polygon + list_overlapping_blue_polygonfor id1 in list_overlapping_all:is_found = Falsefor _, _, _, _, _, bbox_id in list_bboxs:if bbox_id == id1:is_found = Trueif not is_found:# 如果没找到，删除idif id1 in list_overlapping_yellow_polygon:list_overlapping_yellow_polygon.remove(id1)if id1 in list_overlapping_blue_polygon:list_overlapping_blue_polygon.remove(id1)list_overlapping_all.clear()# 清空listlist_bboxs.clear()else:# 如果图像中没有任何的bbox，则清空listlist_overlapping_blue_polygon.clear()list_overlapping_yellow_polygon.clear()# 输出计数信息text_draw = 'DOWN: ' + str(down_count) + \' , UP: ' + str(up_count)output_image_frame = cv2.putText(img=output_image_frame, text=text_draw,org=draw_text_postion,fontFace=font_draw_number,fontScale=0.75, color=(0, 0, 255), thickness=2)cv2.imshow('Counting Demo', output_image_frame)cv2.waitKey(1)capture.release()cv2.destroyAllWindows()

若需要更改模型，只需要更改objdetector.py下面的给出的部分：

OBJ_LIST = ['person', 'car', 'bus', 'truck']
DETECTOR_PATH = 'weights/yolov5m.pt'

总结

本篇文章给出了基于yolov5与DeepSort的流量统计与轨迹跟踪的实例，源码地址：
https://github.com/JulyLi2019/yolov5-deepsort/releases/tag/V1.0