基于DSmT理论的多视角融合目标检测识别

Multi-Perspective Fusion for Object Detection and Recognition Based on DSmT

现有无人车在目标检测中大多依靠单一检测视角进行多传感器数据融合,受传感器检测范围的局限,难以大幅提高准确率,且对融合过程中的类别判定的高冲突情况处理较少.针对以上问题,本文基于多假设思想提出了多视角检测结果的聚类合并方法,并基于DSmT(Dezert-Samarandache theory)和时序信息,改进了冲突分配准则,降低了目标检测的漏检率与误检率.首先利用图像检测算法检测图像中的有效目标,将激光雷达的目标检测结果投影在图像平面上,通过交并比关系构建2种传感器检测结果之间的关联概率矩阵,基于多假设思想实现聚类合并,获取单帧融合检测结果.针对融合过程中可能出现的类别判定冲突情况,利用DSmT融合识别置信度,并结合时序信息对冲突重新分配,获取目标类别的准确识别结果.最后,通过实车实验对算法的有效性进行了验证.

The existing multi-sensor fusion methods for object detection on unmanned vehicles, are mainly based on a single detection perspective. Limited by the scope of the sensor detection, it is difficult for these fusion methods to achieve high accuracy, and deal with the high conflict situation in classification when fusing the results of multiple sensors. To solve the problem and reduce the missing detection rate and false detection rate in object detection, a clustering and combination method of multi-perspective detection results is proposed based on multi-hypothesis theory, and the conflict distribution criterion is improved based on Dezert-Smarandache theory(DSmT) and time sequence information. Firstly, the image detection algorithm is applied to detecting the effective targets in the image, and the detection results of lidar are projected on the image plane. The correlation probability matrix between two kinds of detection results from different sensors is constructed by intersection-over-union. Then the cluster merging is realized based on multiple hypotheses, and the detection result based on single frame fusion is obtained. In order to handle the classification conflict situations in the fusion process, the DSmT is used to fuse the confidence degree, and the conflict is redistributed according to the time sequence information to accomplish accurate object classification. Finally, the validity of the algorithm is verified by real traffic experiment.