摘要
针对复杂交通场景下密集车辆检测存在的目标遮挡、尺度变化大等难题,提出一种基于改进RetinaNet的车辆检测算法SN-RetinaNet。该方法首先在特征提取网络中引入可切换空洞卷积模块,通过动态调整感受野增强多尺度特征提取能力;其次结合神经架构搜索技术优化特征金字塔网络结构,提升算法对不同尺度目标的适应性;最后提出一种基于统计先验的锚框比例优化策略。在SODA10M数据集上的试验结果表明,此方法平均检测精度(mAP)达到48.7%,较基准方法提升3.7个百分点。研究结果为智能交通系统中的车辆检测任务提供了有效的解决方案。
To address the challenges of vehicle occlusion and significant scale variations in dense vehicle detection within complex traffic scenarios,an enhanced RetinaNet-based detection algorithm named SN-RetinaNet is proposed.The method first introduces a switchable atrous convolution module into the feature extraction network,dynamically adjusting the receptive field to enhance multi-scale feature representation.Subsequent neural architecture search is incorporated to optimize the feature pyramid network,improving the model adaptability to objects of varying scales.Finally,a statistical prior-based anchor ratio optimization strategy is proposed.Experiment results on the SODA10M dataset demonstrate that the proposed method achieves a mean average precision(mAP)of 48.7%,outperforming the baseline by 3%.This study provides an effective solution for vehicle detection in intelligent transportation systems.
作者
陈鑫影
吕硕
胡明捷
CHEN Xinying;LYU Shuo;HU Mingjie(School of Rail Intelligence Engineering,Dalian Jiaotong University,Dalian 116028,China)
出处
《大连交通大学学报》
2025年第4期147-154,共8页
Journal of Dalian Jiaotong University
基金
辽宁省应用基础研究计划项目(1655706734383)。
关键词
目标检测
可切换空洞卷积
特征金字塔网络
神经架构搜索
target detection
switchable atrous convolution
feature pyramid networks
neural architecture search
