基于YOLOv5和DeepSort算法的工程车辆识别与多目标跟踪实现  

Implementation of Engineering Vehicle Recognition and Multi-Object Tracking Based on YOLOv5 and DeepSORT Algorithms

作　　者：孙长虹 孙洪亮 李轩 Sun Changhong;Sun Hongliang;Li Xuan(Heilongjiang Transportation Investment Engineering Construction Co.,Ltd.,Harbin,China)

机构地区：[1]黑龙江省交投工程建设有限公司,黑龙江哈尔滨

出　　处：《科学技术创新》2025年第4期111-114,共4页Scientific and Technological Innovation

摘　　要：本研究聚焦于传统目标检测技术与基于深度神经网络的工程车辆检测策略的对比分析。通过借助有效技术手段,采用降噪、增强及边缘检测的方式,对图像的质量进行有效优化。为了确保工程车辆检测过程的专业性与高精度,我们借助YOLOv5算法对其进行了处理,该算法的运用可以进一步提高处理速度。对于检测时容易出现的目标遗漏与预测框定位不准确情况,我们借助DeepSORT算法,通过全面的整合对检测目标进行了追踪预测。DeepSORT通过卡尔曼滤波进行数据估计,能实现高效的连续跟踪。为应对拍摄设备晃动及车辆变速行驶引发的目标身份频繁更迭挑战,我们创新性地采用了一种改进的GIoU计算方法。This study focuses on the comparative analysis of conventional target detection techniques and engineering vehicle detection strategies based on deep neural networks.Optimize the image quality by means of noise reduction,enhancement and edge detection.In order to ensure the professionalism and high precision of the engineering vehicle detection process,we have processed it with the help of YOLOv5 algorithm,which can further improve the processing speed.For the cases of target omission and inaccurate prediction box positioning that are easy to occur during detection,we tracked and predicted the detected target through comprehensive integration with the help of DeepSORT algorithm.DeepSORT Data estimation by Kalman filter can achieve efficient continuous tracking.In order to cope with the challenge of frequent change of target identity caused by the shaking of camera equipment and vehicle speed,we innovatively adopted an improved GIoU calculation method.

关 键 词：YOLOv5算法 工程车辆检测 DeepSORT算法 多目标跟踪 实时检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]