基于交并比匹配的实时决策级融合检测框架  

Real-time Decision-level Fusion Detection Framework Based on Intersection over Union Matching

作　　者：苏向阳 汪洋 谭森起 张乃斯 罗天 荣志刚 SU Xiangyang;WANG Yang;TAN Senqi;ZHANG Naisi;LUO Tian;RONG Zhigang(China North Artificial Intelligence&Innovation Research Institute,Beijing 100072,China;Collective Intelligence&Collaboration Laboratory,Beijing 100072,China;China North Vehicle Research Institute,Beijing 100072,China)

机构地区：[1]中兵智能创新研究院有限公司,北京100072 [2]群体协同与自主实验室,北京100072 [3]中国北方车辆研究所,北京100072

出　　处：《无人系统技术》2025年第1期41-49,共9页Unmanned Systems Technology

基　　金：国家自然科学基金(52202512)。

摘　　要：随着无人驾驶技术的快速发展,精确的行人检测已成为环境感知中不可或缺的关键环节。针对现有融合框架训练成本高、实时性差导致难以在无人车上部署的问题,提出了一种基于交并比匹配的实时决策级融合检测框架。首先,以高实时性的YOLO-P模型作为主检测器,利用摄像头和激光雷达之间的位姿关系,提取点云行人感兴趣区域以减少计算量。其次,提出了一种适应行人行为的聚类方法,对感兴趣区域的障碍物进行聚类。然后,将障碍物的三维边界框投影至图像坐标系,求取与图像行人检测二维边界框的交并比,进行障碍物与行人的匹配。最后,提出了置信度修正函数,对较低置信度的行人检测结果进行修正,提高行人检测准确率。实验结果表明,该融合检测策略在小间隔、较远距离及遮挡场景下的检测效果良好,平均准确率达到了95.38%,同时,检测帧率为29.2 FPS,满足无人驾驶对实时性的要求。该融合检测框架有望在更广泛的无人驾驶场景中得到应用,进一步提升无人驾驶系统的安全性和可靠性。With the rapid development of autonomous driving technology,accurate pedestrian detection has become an indispensable key component in environmental perception.To address the issues of high training costs and poor real-time performance in existing fusion frameworks,which hinder deployment on autonomous vehicles,a real-time decision-level fusion detection framework based on Intersection over Union(IoU)matching is proposed.Firstly,the high real-time YOLO-P model is used as the main detector.By utilizing the calibration relationship between the camera and the LiDAR,the region of interest for pedestrians in the point cloud is extracted to reduce the computational load.Secondly,a clustering method adapted to pedestrian behavior is proposed to cluster obstacles in the region of interest.Then,the 3D bounding boxes of obstacles are projected into the image coordinate system,and the IoU with the 2D bounding boxes from pedestrian detection in the image is calculated to match obstacles with pedestrians.Finally,a confidence correction function is introduced to refine the pedestrian detection results with low confidence,thereby enhancing the accuracy of pedestrian detection.The experimental results show that the fusion detection strategy performs well in scenarios with small intervals,long distances,and occlusions.The average accuracy reaches 95.38%,and the detection frame rate is 29.2 FPS,which meets the real-time requirements of autonomous driving.This fusion detection framework is expected to be applied in a wider range of autonomous driving scenarios,further enhancing the safety and reliability of autonomous driving systems.

关 键 词：多传感器融合 YOLO-P模型 交并比匹配 实时检测 无人车 点云聚类 联合标定 

分 类 号：U461[机械工程—车辆工程]