1. 动机
目标跟踪是计算机视觉领域一种常用的算法，用于将前后帧中的同一个目标关联起来，从而可以针对某一个特定目标进行分析，如对状态进行投票平滑获取更为稳健的结果。
然而，目前流行的跟踪算法大多是基于检测的bbox之间的IOU来匹配的，这对于某些小目标或者点的检测，IOU通常不是一个好的选择，因为目标太小，很容易使得相邻两帧之间的IOU为0。
为了解决这个问题，本文提出了一种基于点之间距离的跟踪方法：将目标建模为一个点，通过计算前后帧点之间的距离，利用匈牙利匹配来进行跟踪。
2. 方法
直接上代码，里面给出了跟踪方法的定义以及一个使用示例：
"""
Test for multi-target tracker
"""import numpy as np
from scipy.optimize import linear_sum_assignment# 定义目标类
class Target:def __init__(self, x, y):self.x = xself.y = yself.id = Noneself.miss_num = 0self.lost = False # 添加一个标记来表示目标是否丢失# 定义跟踪器类
class Tracker:def __init__(self, max_age=1, thres_dist=100):self.max_age = max_ageself.thres_dist = thres_distself.tracked_targets = []self.last_id = 0# 在下一帧中更新目标位置并使用匈牙利算法进行匹配def update(self, targets):# print([[t.x, t.y] for t in targets])# print([[t.x, t.y] for t in self.tracked_targets])# 计算每个跟踪器与目标之间的相似度（这里使用欧氏距离作为相似度指标）distances = []for target in targets:dist_list = []for tracked_target in self.tracked_targets:if not tracked_target.lost: # 只考虑未丢失的目标dist = np.sqrt((target[0] - tracked_target.x) ** 2 + (target[1] - tracked_target.y) ** 2)dist_list.append(dist)else:dist_list.append(np.inf) # 将丢失的目标设置为无穷大距离，避免被匹配distances.append(dist_list)distances = np.asarray(distances)# 使用匈牙利算法进行匹配row_ind, col_ind = linear_sum_assignment(distances)# print(row_ind, col_ind, distances)# 更新匹配成功的目标位置# print(len(self.tracked_targets))for i, j in zip(row_ind, col_ind):if not self.tracked_targets[j].lost and distances[i][j] < self.thres_dist: # 只更新未丢失的目标位置self.tracked_targets[j].x = targets[i][0]self.tracked_targets[j].y = targets[i][1]else:# 如果目标丢失，继续标记为丢失状态，不进行位置更新，同时累计丢失次数self.tracked_targets[j].miss_num += 1if self.tracked_targets[j].miss_num >= self.max_age:# print("lost", self.tracked_targets[j].id)self.tracked_targets[j].lost = True# 添加新的目标到跟踪列表for j in range(len(targets)):if j not in col_ind:self.create_target(targets[j][0], targets[j][1])# 检测丢失的目标，如果目标丢失超过一定帧数，将其从跟踪列表中删除lost_targets = []for target in self.tracked_targets:if target.lost:lost_targets.append(target)for target in lost_targets:self.tracked_targets.remove(target)def create_target(self, x, y):tar = Target(x, y)tar.id = self.last_idself.tracked_targets.append(tar)self.last_id += 1def get_color_by_id(id):colors = [(255, 0, 0),(255, 255, 0),(255, 0, 255),(30, 140, 100),(0, 255, 0),(0, 50, 200),(100, 0, 30),(100, 100, 0),(20, 10, 200),(20, 250, 100),(145, 0, 90),(15, 10, 190),(15, 100, 100),]index = id % len(colors)return colors[index]if __name__ == '__main__':import cv2from PIL import Imageimport copyimg = np.ones([500, 500, 3]).astype(np.uint8)pil_imgs = [Image.fromarray(img)]# 初始化目标列表preds = [[1, 1], [1, 499], [499, 1], [499, 499]]# 初始化跟踪器tracker = Tracker()# 模拟多帧跟踪过程for i in range(100):# 在每一帧中，随机移动每个目标的位置delta = 10if i in [10, 11, 12, 13]:delta = 30# print([[t.x, t.y] for t in tracker.tracked_targets])preds[0][0] += np.random.randint(0, delta)preds[0][1] += np.random.randint(0, delta)preds[1][0] += np.random.randint(0, delta)preds[1][1] -= np.random.randint(0, delta)preds[2][0] -= np.random.randint(0, delta)preds[2][1] += np.random.randint(0, delta)preds[3][0] -= np.random.randint(0, delta)preds[3][1] -= np.random.randint(0, delta)# print([[t.x, t.y] for t in tracker.tracked_targets])# 更新跟踪器的目标列表tracker.update(copy.deepcopy(preds))# 打印当前帧中每个跟踪到的目标位置# print("frame: {}".format(i))points, ids = [], []for target in tracker.tracked_targets:print("Target {}: ({}, {}); lost: {}".format(target.id, target.x, target.y, target.lost))points.append([target.x, target.y])ids.append(target.id)# cv2.circle(img, [target.x, target.y], point_size, point_color, thickness)cv2.putText(img, str(target.id), [target.x, target.y], cv2.FONT_HERSHEY_SIMPLEX, 0.4,get_color_by_id(target.id), thickness=1, lineType=cv2.LINE_AA)pil_imgs.append(Image.fromarray(img))cv2.imshow('img', img)if cv2.waitKey(30) == ord('q'):break# 创建并保存GIF文件image_0 = pil_imgs[0]image_0.save('track_result.gif', save_all=True, append_images=pil_imgs[1:], duration=30, loop=0)
运行上述代码后，会保存一个GIF文件，展示了多个目标的跟踪结果，如下图：
