基于融合差分卷积的受电弓安全触发目标实时检测定位方法  

Pantograph Safe Trigger Target Real-Time Detection and Localization Method Based on Fused Differential Convolutional

作　　者：杨占山 张瀛[1] 杜弘志 孙岩标 邾继贵[1] Yang Zhanshan;Zhang Ying;Du Hongzhi;Sun Yanbiao;Zhu Jigui(State Key Laboratory of Precision Measuring Technology and Instruments,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学精密测试技术及仪器全国重点实验室,天津300072

出　　处：《激光与光电子学进展》2024年第18期77-85,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(52075382)。

摘　　要：针对现有目标检测算法存在的问题,提出了一种基于融合差分卷积的目标实时检测定位方法。首先构建融合差分卷积的主干网络以增强特征提取能力;然后设计共享权重的特征融合模块和检测头以提高检测速度和精度;最后制定多阶段训练策略进一步提升精度。在受电弓检测数据集中的实验结果表明,在CPU硬件资源下,所提方法检测帧率可达149 frame/s,整体平均精度均值(mAP)可达81.20%,比FemtoDet算法分别提高了57 frame/s和6百分点。所提方法满足高速铁路现场中对触发定位任务的实时性和准确性的技术需求。Aiming at the problems of existing target detection algorithms,a real-time target detection and localization method based on fused differential convolution is proposed.Firstly,a backbone network with fused differential convolution is constructed to enhance feature extraction capabilities.Then,a feature fusion module and detection head with shared weights are designed to improve detection speed and accuracy.Finally,a multi-stage training strategy is formulated to further enhance accuracy.Experimental results on the pantograph detection dataset show that the proposed method achieves a frame detection speed of up to 149 frame/s on CPU hardware resources,with an whole mean average precision(mAP)of 81.20%.This is an improvement of 57 frame/s and 6 percentage points compared to the FemtoDet algorithm.Proposed method meets technical requirements for real-time and accurate triggering positioning tasks in high-speed railway scenarios.

关 键 词：目标检测 模型压缩 特征融合 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]