随着城市轨道交通规模的日益扩大,异物侵限已成为城市轨道交通安全的重大隐患,基于人工智能方法开展城市轨道交通异物检测成为研究热点.相对于传统视觉相机,激光雷达具有光照不敏感、测距精度高、可远距离探测等优点,更适用于城市轨道交通场景下的安全监控和检测.面对海量的稀疏点云数据,现有基于激光雷达的目标检测方法会因三维结构信息损失产生漏检、误检问题.针对以上不足,提出一种基于结构密度感知的单阶段目标检测方法(Structure Density Aware Single-Stage Object Detector,SDA-SSD).设计体素特征聚合模块提取三维结构信息,通过三重特征融合模块将点云特征和三维结构信息融合,改善因提取高层语义而导致空间特征质量下降的问题,提升目标的检测能力.设计体素密度值用于度量样本的稀疏程度,基于体素密度值校正分类置信度,改善目标定位和分类精度不一致的问题.实验结果表明:所提算法在KITTI数据集汽车类别的平均检测精度为88.21%,检测速度为21 FPS,相较于基准网络SECOND,平均检测精度提高了2.36个百分点,检测速度提高了13%,具有较高的识别率和实时性.在城市轨道交通实际场景对所提算法进行了验证,算法在复杂场景下具有良好的检测效果,能准确识别到列车前方目标障碍物,具备较高的有效性和可行性.研究成果为城市轨道交通安全营运和人民生命财产安全提供了保障.
Rail foreign object detection based on SDA-SSD
With the increasing scale of urban rail transportation,foreign object encroachment has become a major potential hazards for transit safety,and the foreign objects detection based on artificial intelligence methods has become a research hotspot.Compared with traditional vision camera and LiDAR methods,those methods based on artificial intelligence have the advantages of light insensitivity,high ranging accuracy,and long-distance detection,which are more suitable for safety monitoring and detection in rail transportation scenarios.In face of massive point cloud data,the existing LiDAR-based target detection methods may lead to leakage and false detection results due to the loss of 3D structure information.To address the above problem,a single-stage target detection method SDA-SSD(Structure Density Aware Single-Stage Object Detector)based on structure density awareness was proposed.The voxel feature aggregation module was designed to extract 3D structure information,and the triple feature fusion module was designed to fuse point cloud features and 3D structure information to prevent the spatial feature quality degradation caused by the extraction of high-level semantics and enhance the detection performance.The voxel density value was introduced to measure the sparsity of samples,and the classification confidence was corrected based on the voxel density value to improve the performance of inconsistent target localization and classification.The experimental results show that the proposed method achieves 88.21%average accuracy and 21 FPS detection speed in the car category of KITTI dataset,which can improve the average detection accuracy by 2.36 percentage points and detection speed by 13%compared with the benchmark network SECOND.Moreover,the proposed method was verified against the dataset in which data is collected in the actual scenario of urban rail transit.The target obstacles in front of the train can be detected successfully.Finally,the proposed method has satisfactory detection performance in the complex scenarios and can improve the safety of the rail transit effectively.
国家科技期刊平台
NSTL
万方数据
