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

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.