一种抗遮挡重叠与尺度变化的行人检测算法OA

In view of the facts that the detection accuracy of pedestrian detector decreases and the missed detection rate increases due to factors such as the occlusion and the different scales of pedestrians(the objects)in complex crowded scenes,an improved pedestrian detection algorithm that is resistant to occlusion overlap and scale variation is proposed based on Faster R-CNN algorithm.In the step of feature extraction,a recurrent multi-scale feature extraction network is designed to incorporate the attention mechanism,which is used to learn more rich and detailed multi-scale feature information and focus on the key feature information to improve the sensitivity of the network to pedestrians(the objects)of different scales.For the loss function module,repulsive loss is introduced to reduce the interference(caused by mutual occlusion of the objects)on detection.In the step of postprocessing,a non-maximum suppression(NMS)algorithm based on the compensation for overlapping rate is designed,so that the actual suppression threshold can be adaptively adjusted with the change of the degree of occlusion,thus further reducing the missed detection of the pedestrians(the objects)in dense places.Experimental results show that the improved algorithm has better detection performance.Its average detection accuracy on the CrowdHuman dataset and CityPersons dataset is improved by 2.5%and 1.9%,respectively,and its log-average miss rates(LAMRs)are reduced by 3.5%and 3.2%,respectively,in comparison with those of the baseline algorithm.In addition,its LAMR of pedestrians(the objects)of different scales on the TJU-DHD-pedestrian dataset is also reduced significantly.The proposed algorithm can be applied to pedestrian detection in complex scenes.