基于深层聚合网络的车辆多目标检测与追踪  

Vehicle multi-object detection and tracking algorithmbased on deep aggregation network

作　　者：郭俊麟 董超俊 陆晓田 GUO Jun-lin;DONG Chao-jun;LU Xiao-tian(Department of Intelligent Manufacturing,Wuyi University,Jiangmen 529000,China)

机构地区：[1]五邑大学智能制造学部,广东江门529000

出　　处：《计算机工程与设计》2023年第7期2080-2086,共7页Computer Engineering and Design

基　　金：广东省省级科技计划基金项目(2017A010101019)。

摘　　要：针对现有车辆多目标检测与追踪算法在长时间遮挡、光线变化等因素的影响下造成的ID切换和误检问题,提出一种基于深层聚合网络的改进算法,进行车辆检测并完成特征嵌入任务。通过在下采样过程中融合注意力模块进行特征增强,在特征融合过程中使用可变形卷积提高捕获不同尺度目标的能力,采用高斯核函数在目标中心区域提取车辆特征信息用于轨迹关联,实现车辆多目标追踪任务。所提算法在UA-DETRAC数据集上进行实验,其MOTA和IDF1分别达到了79.37%和88.12%,ID切换次数从77下降至36,满足实时检测与追踪的需求。Aiming at the problems of ID switch and false detection in current vehicle multi-object detection and tracking algorithm caused by the long-term occlusion,illumination variation or other factors,an improved algorithm based on deep aggregation network was proposed to detect vehicle and complete the feature embedding tasks.The feature enhancement was performed by integrating the attention module in the downsampling process.The deformable convolution was used in the feature fusion process to improve the ability to capture targets of different scales.The Gaussian kernel function was used to extract vehicle feature information in the target’s center region for trajectory association to realize the multi-object track task of the vehicle.The proposed algorithm was experimented on the UA-DETRAC dataset.The MOTA and IDF1 of the algorithm reach 79.37%and 88.12%respectively,the number of ID switching decreases from 77 to 36,and the needs of real-time detection and tracking are met.

关 键 词：目标检测 多目标追踪 深层聚合网络 特征嵌入 多特征融合 注意力机制 数据关联 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]