A Pedestrian Detection Model Based on Binocular Information Fusion

Abstract

Pedestrian detection, as a special kind of target detection, is a research hotspot in the field of image processing and computer vision. Because monocular vision cannot obtain the depth information of the image, it cannot meet the accuracy requirements of pedestrian detection. In order to solve these problems, a new cascading pedestrian detection model based on PSMNet binocular information fusion and improved faster R-CNN pedestrian detection model is proposed. Firstly, binocular images are fed into the original PSMNet binocular information fusion module to get the disparity map, and then left and right images are fused by the disparity map to get the fusion image. Secondly, in the improved faster R-CNN pedestrian detection module, the left, right and fusion image of one frame are as separate inputs, and pedestrian detection is carried out respectively. Finally, the detection results of the three channels are passed through the target consistency validation module, and the verified pedestrian detection target is as the final output detection result. The simulation results show that the accuracy and recall rate of the cascading model are improved, the missed detection rate is reduced to 13.42%, and the accuracy rate is 88.58%.