基于视锥距离和自适应权重卡尔曼滤波的多传感器融合算法  

A Multi-sensor Fusion Algorithm Based on View-cone Distance and Adaptive Weighted Kalman Filter

作　　者：李杰 张洛维 王晓燕[2] 胡铮[3] 兰海[3] 王志勇[3] 王莉 LI Jie;ZHANG Luo-wei;WANG Xiao-yan;HU Zheng;LAN Hai;WANG Zhi-yong;WANG Li(l.School of Mechanical-electronic and Automobile Engineering,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;School of Information,Beijing Wuzi University,Beijing 101149,China;Key Lab of Vehicular Transmission,China North Vehicle Research Institute,Beijing 100072,China)

机构地区：[1]北京建筑大学机电与车辆工程学院,北京102616 [2]北京物资学院信息学院,北京101149 [3]中国北方车辆研究所车辆传动重点实验室,北京100072

出　　处：《中国公路学报》2024年第3期194-203,共10页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(51675494);北京建筑大学金字塔人才培养工程项目(JDJQ20200308)。

摘　　要：为提升智能驾驶系统的环境感知能力,采用多模态传感器并结合人工智能技术解决单模态传感器在环境感知方面存在识别效果差、易受干扰等问题。然而,跨模态传感器间的特征匹配仍存在问题,如特征表示不一致、感知误差、延迟误差等。为解决这些问题,提出一种基于视锥距离度量的方法,构建了目标匹配矩阵,使用匈牙利算法进行帧间关联匹配。并引入自适应权重调节技术优化卡尔曼滤波算法,实现了低复杂度高效的跨模态传感器融合目标检测与跟踪。对比交并比(IoU)度量和欧氏距离度量,所提方法在骑行者和行人类别中,多目标跟踪准确度(MOTA)分别提升至81.22%和58.62%。研究结果表明,融合方法的均方根误差达到了0.349 0,相比相机和激光雷达单独预测方法分别减少了30.37%和30.53%,证实了所提出的自适应权重卡尔曼滤波融合跟踪方法的准确性。在KITTI数据集上的多目标跟踪实验测试,准确度达到了88.25%,与当前主流方法性能相当。在多种天气环境下的测试结果也展示了优异的性能,车辆、行人、骑行者的目标检测准确率分别达到了96.40%、75.51%和91.87%。相较于单一传感器,该融合方法在多种路况下表现出更优越的检测效果,提高了系统的可靠性与鲁棒性,为无人驾驶技术的进一步发展奠定了坚实基础。To improve the environmental perception ability of intelligent driving systems,multimodal sensors have been integrated with the artificial intelligence technology to solve the problem of single-modal sensors,such as poor recognition and vulnerability to interference in environmental perception.However,problems of feature matching between cross-modal sensors,such as inconsistent feature representations,sensing errors,and delay errors,persist.To address these problems,thisstudy proposes a view-cone distance metric-based method,constructs a target-matchingg matrix,and uses the Hungarian algorithm for inter-frame association matching.Based on the intersection over union(IOU)and Euclidean distance metrics,the multiple object tracking accuracy(MOTA)of the proposed method improved to 81.22%and 58.62%in the cyclist and pedestrian categories,respectively.An adaptive weightadjustment technique was also introduced to optimize the Kalman filter algorithm so that lowcomplexity and efficient cross-modal sensor fusion target detection and tracking could be achieved.Compared with the individual camera and light detection and ranging(LIDAR)predictions,the root mean square error of the fusion method reaches 0.3490,denoting a reduction by 30.37%and 30.53%compared with the camera and LIDAR methods,respectively,and confirming the accuracy of the proposed adaptive weighting Kalman filter fusion tracking method.The multi-target tracking experimental test conducted on the KITTI dataset achieved an accuracy of 88.25%,comparable to the performance of current mainstream methods.The test results under multiple weather conditions also demonstrated excellent performance,with target detection accuracies of 96.40%,75.51%,and 91.87%for vehicles,pedestrians,and cyclists,respectively.Compared to a single sensor,the fusion method attained superior detection results under multiple road conditions,improved the reliability and robustness of the system,and laid a solid foundation for the further development of driverless technology.

关 键 词：汽车工程 多目标跟踪 传感器融合 视锥距离 卡尔曼滤波 

分 类 号：U469.79[机械工程—车辆工程]