摘要
针对基于深度学习的目标跟踪算法实时性差、成功率低以及抗干扰性弱等问题,提出了一种基于改进YOLOv3和DSST的目标跟踪算法。选用检测精度和速度均优于SDD的YOLOv3检测算法,并通过减少YOLOv3算法中1个尺度的输出张量以提高检测实时性;将DSST跟踪算法预测的目标区域放大2倍后作为YOLOv3检测算法输入,检测结果用于更新DSST跟踪目标框,从而提高跟踪算法的抗干扰性。实验表明:提出的算法提高了跟踪算法成功率和实时性,在很多场景下表现出较强的鲁棒性。
Aiming at the problems of poor real-time performance, low success rate and weak anti-jamming of target tracking algorithm based on deep learning, a target tracking algorithm based on improved YOLOv3 and DSST is proposed. The YOLOv3 detection algorithm with better detection accuracy and speed than SDD was selected, and the real-time detection performance was improved by reducing the output tensor of one scale in YOLOv3 algorithm. The target area predicted with DSST tracking algorithm was enlarged twice and then used as the input of YOLOv3 detection algorithm. The detection result was used to update the target frame of DSST tracking, so as to improve the anti-interference performance of tracking algorithm. Experiments show that the proposed algorithm improves the tracking algorithm’s success rate and real-time performance, and strong robustness in many scenarios is available.
作者
蔡锦华
祝义荣
CAI Jin-hua;ZHU Yi-rong(Jiangxi Lianchuang Precision Electromechanical C o.,Ltd,Nanchang Jiangxi 330096,China)
出处
《计算机仿真》
北大核心
2020年第5期213-217,321,共6页
Computer Simulation
关键词
目标跟踪
算法
网络
Target Tracking
Algorithm
Network
