基于改进YOLOv3的多目标跟踪算法研究

Multitarget Tracking Algorithm Based on an Improved YOLOv3 Algorithm

针对目前多目标跟踪过程中漏检率高和检测速率慢的问题,提出一种改进YOLOv3网络结构的多目标跟踪算法。首先,利用K-means++聚类算法对数据集中的目标边框进行聚类,根据聚类结果优化网络的先验框参数。然后,在Darknet-53特征提取层中引入深度可分离卷积模块,用深度可分离卷积代替标准卷积,减少参数量,并在YOLO预测层中引入SENet模块,利用SENet模块突出特征图的关键通道信息。最后,选定经典的trackingby-detection框架,使用改进的YOLOv3算法来实现对目标信息的检测工作,跟踪部分选用Deep-SORT算法进行跟踪。实验结果表明,所提出的多目标跟踪算法能够有效地减小漏检率,同时兼顾了算法的检测精度和实时性。

To solve the problem of high missed rate and slow detection rate in the current multitarget tracking process,a multitarget tracking algorithm with an improved YOLOv3 network structure is proposed.First,the K-means++algorithm is utilized to cluster the target boundaries in the dataset.The priori parameters of the network are optimized using the clustering results.Then,the deep separable convolution module is employed instead of standard convolution in the Darknet-53 feature extraction layer,thereby reducing the number of parameters.In addition,the key channel information of the feature map is highlighted by applying the SENet module in the YOLO prediction layer.Finally,the improved YOLOv3 algorithm is used to implement the detection of a target in the classic tracking-by-detection framework.Meanwhile,the Deep-SORT algorithm is adopted in the tracking part.Experimental results show that the proposed multitarget tracking algorithm can effectively reduce the missed detection rate and take into account the detection accuracy and real-time performance,simultaneously.