摘要
针对密集交通场景中车辆跟踪所面临的漏检率高、误检频发以及跟踪精度低等问题,提出一种基于DeepSort多目标车辆跟踪优化算法,旨在提升其在复杂环境下的跟踪性能。首先,优化卡尔曼滤波,通过增加自适应调制噪声尺度的机制,动态调整噪声协方差,更准确地预测目标的运动轨迹,克服因噪声水平波动导致的预测偏差和不稳定现象。随后,采用ResNest50作为主干网络,并结合YOLOv5检测器,对外观特征提取网络进行了改进,增强对车辆外观特征的精细提取能力,达到准确检测跟踪场景中的多个目标车辆的目的。为了全面评估所提算法的性能,基于MOTA、IDSW和MOTP三种指标,进行了综合评估实验并辅以可视化对比实验。经过实验验证,所提算法与其他模型相比在MOTA、MOTP上平均提升了3.460%、3.103%,同时使IDSW平均下降了45.907%,提高了多目标车辆检测的鲁棒性及跟踪精度。
Aiming at the problems of high miss detection rate,frequent false detection and low tracking accuracy in vehicle tracking in dense traffic scenes,an optimization algorithm for multi-target vehicle tracking based on DeepSort is proposed to improve its tracking performance in complex environments.Firstly,Kalman filter is optimized to dynamically adjust noise covariance by adding adaptive modulation noise scale mechanism,so as to predict target trajectory more accurately and overcome prediction bias and instability caused by noise level fluctuation.Then,ResNest50 is used as the backbone network,combined with YOLOv5 detector,the appearance feature extraction network is improved to enhance the fine extraction ability of vehicle appearance features,so as to accurately detect multiple target vehicles in the tracking scene.In order to evaluate the performance of the proposed algorithm comprehensively,a comprehensive evaluation experiment is carried out based on MOTA,IDSW and MOTP,supplemented by visual contrast experiments.Compared with other models,the proposed algorithm improves MOTA and MOTP by 3.460%and 3.103%respectively,and reduces IDSW by 45.907%,which improves the robustness and tracking accuracy of multi-target vehicle detection.
作者
余海燕
霍爱清
冯若水
YU Haiyan;HUO Aiqing;FENG Ruoshui(School of Electronic Engineering,Xi′an Shiyou University,Xi′an 710065,China)
出处
《智能计算机与应用》
2025年第5期68-74,共7页
Intelligent Computer and Applications
基金
陕西省科技厅一般项目(2020GY-152)。
