基于联合检测的多目标跟踪方法研究  

Research on multitarget tracking method based on joint detection

作　　者：郭文杰 聂国豪 王兴梅[1,2] 赵一霖 GUO Wenjie;NIE Guohao;WANG Xingmei;ZHAO Yilin(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China;National Key Laboratory of Underwater Acoustic Technology,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001 [2]哈尔滨工程大学水声技术全国重点实验室,黑龙江哈尔滨150001

出　　处：《应用科技》2024年第2期8-16,共9页Applied Science and Technology

基　　金：重点实验室开放基金项目(KY10600220048).

摘　　要：为了更好地应对多目标跟踪联合检测算法面对的场景遮挡问题,通过结合注意力机制,提出基于Transformer的运动预测和数据关联(Transformer-based motion prediction and data association,TrMPDA)联合检测跟踪方法。首先,考虑到置信度检测框的质量以及深度特征的视觉表示能力对遮挡场景下跟踪效果的影响,重新设计TrMPDA骨干网络中的ResNet卷积模块,利用相邻像素和长距离像素间丰富的上下文关系指导动态注意矩阵的学习,增强深度特征的视觉表示能力,并通过边界框的宽和高估计边界框位置,提高置信度检测框的质量。其次,在本文方法中保留所有的检测框,根据阈值大小划分高置信度检测框和低置信度检测框,分别执行数据关联匹配,以此来平衡由于遮挡导致的检测框低置信度。实验结果表明本文提出的TrMPDA方法与典型的Sort、JDE、Fairmot等多目标跟踪算法相比具有更好的跟踪效果,能够应对多目标跟踪中目标遮挡的问题。Recent tracking methods based on joint detection have demonstrated high performance in multi-object tracking.However,these algorithms have limitations in tracking performance in certain scenarios such as occlusion.To address this,the present paper introduces a joint detection tracking method based on Transformer for motion prediction and data association(TrMPDA)by incorporating attention mechanism into the joint detection based multi-object tracking method.Firstly,the ResNet convolution module in the backbone network of TrMPDA was redesigned to account for the influence of the quality of confidence detection boxes and the visual representation ability of deep features in occlusion scenarios.This involved utilizing the contextual relationships between adjacent and long-distance pixels to guide the learning of dynamic attention matrices,thereby enhancing the visual representation ability of deep features,and estimating the position of the bounding box through its width and height to improve the quality of the confidence detection box.Secondly,in the proposed method,all detection boxes were retained,and high-confidence and low-confidence detection boxes were distinguished based on a threshold.Data association matching was separately performed for high-confidence and low-confidence detection boxes to address potential cases of low-confidence detection boxes being influenced by occlusion.Experimental results demonstrate that compared with typical multiobject tracking algorithms such as Sort,JDE,and Fairmot,the proposed TrMPDA method exhibits strong tracking performance,and is effective in coping with the challenge of object occlusion in multi-object tracking.

关 键 词：运动预测 注意力机制 数据关联 卡尔曼滤波 目标遮挡 动态注意矩阵 联合检测 多目标跟踪 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]