基于深度学习和颜色特征的行人跟踪算法

Pedestrian Tracking Algorithm Based on Deep Learning and Color Features

针对行人跟踪算法中因行人遮挡而导致行人跟踪准确率低、跟踪速度慢的问题,论文提出了一种基于深度学习和颜色特征的行人跟踪算法。首先利用yolov5目标检测算法检测行人,得到带有行人框的视频帧,同时利用检测框坐标信息判断行人之间是否存在遮挡,若有遮挡,则把遮挡区域像素设为0,分割出非遮挡区域,将非遮挡区域转化为HSV颜色空间,量化HSV分量,构造颜色特征直方图,并表示为一维向量G。其次,以第一帧行人检测框坐标为基础构建行人跟踪模型,初始化跟踪对象,并根据行人质心变化预测行人位置。在公开数据集MOT-16数据集上测试,MOTA为49.78%,相比于Sort和DeepSort算法分别提高1.51%和0.33%,在IDF1分数上分别高于Sort和DeepSort算法7.07%和3.46%。跟踪速度比DeepSort提升24%。

Aiming at the problems of low pedestrian tracking accuracy and slow tracking speed caused by pedestrian occlusion in the pedestrian tracking algorithm,this paper proposes a pedestrian tracking algorithm based on deep learning and color features.First,it uses the yolov5 target detection algorithm to detect pedestrians,and obtains video frames with pedestrian frames.At the same time,the coordinate information of the detection frame is used to determine whether there is occlusion between pedestrians.If there is occlusion,the pixels of the occlusion area is set to 0,and the non-occlusion area is segmented,the non-occluded area is converted into the HSV color space,the HSV component is quantized,a color feature histogram is constructed,and it is expressed as a one-dimensional vector G.Secondly,the pedestrian tracking model is constructed based on the coordinates of the pedestrian de-tection frame in the first frame,the tracking object is initialized,and the pedestrian position is predicted according to the change of the pedestrian's centroid.Tested on the public data set MOT-16 data set,the MOTA is 49.78%,which is 1.51%and 0.33%higher than the Sort and DeepSort algorithms,respectively,and 7.07%and 3.46%higher than the Sort and DeepSort algorithms in the IDF1 score.The tracking speed is 24%higher than that of DeepSort.