基于改进YOLOv7-DeepSort的红外视频多目标跟踪  

Infrared Video Multi Target Tracking Based on Improved YOLOv7-DeepSort

作　　者：宫华[1,3] 张众垚 胡雨桐 刘芳 GONG Hua;ZHANG Zhongyao;HU Yutong;LIU Fang(Shenyang Ligong University,Shenyang 110159,China;Liaoning Key Laboratory of Intelligent Optimization and Control for Ordnance Industry,shenyang 110159,China;National Key Laboratory of Electromagnetic Space Security,Tianjin 300308,China)

机构地区：[1]沈阳理工大学理学院,沈阳110159 [2]沈阳理工大学自动化与电气工程学院,沈阳110159 [3]辽宁省兵器工业智能优化与控制重点实验室,沈阳110159 [4]电磁空间安全国家重点实验室,天津300308

出　　处：《沈阳理工大学学报》2024年第6期20-27,共8页Journal of Shenyang Ligong University

基　　金：军委科技委重点实验室预研基金项目(2022JCJQLB055008)。

摘　　要：针对红外图像纹理弱及多目标遮挡导致跟踪精度低的问题,构建了基于改进YOLOv7模型和多目标跟踪算法DeepSort的融合红外目标跟踪模型MSB-YOLOv7-DeepSort。采用SE(squeeze and excitation)通道注意力机制和双向特征金字塔网络提高红外目标的特征提取质量;利用轻量化网络MobileNetV3替换YOLOv7骨干网络,提升融合模型的推理速度。实验结果表明,MSB-YOLOv7-DeepSort模型在跟踪准确度、跟踪精确度、正确目标跟踪比例和帧率等方面均具有较好的性能。To solve the problem of low tracking accuracy caused by weak texture and multi target occlusion,a fusion infrared tracking model MSB-YOLOv7-DeepSort is constructed,based on the improved YOLOv7 and the multi-target tracking algorithm DeepSort.The SE(squeeze and excita-tion)channel attention mechanism and the bidirectional feature pyramid network are utilized to im-prove the quality of feature extraction for the infrared targets.The lightweight network Mobile-NetV3 is used to replace the YOLOv7 backbone network to enhance the inference speed of the fu-sion model.The experimental results indicate that the MSB-YOLOv7-DeepSort model has good per-formance in tracking accuracy and frame rate.

关 键 词：红外目标跟踪 YOLOv7 轻量化 SE注意力机制 MobileNetV3 双向特征金字塔网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]