| Pedestrian detection is one of the hot research directions in target detection,and its purpose is to detect the location of the target pedestrian from a static image or video sequence through the comprehensive use of various technologi es.With the increasing computing power of computers in recent years,pedestrian detection algorithms based on convolutional neural networks have made countless breakthrough advances.Although recent technological developments have greatly improved pedestrian detection,there is still much room for improvement in the current stage of pedestrian detection models.This paper focuses on the multi-scale and occlusion problems faced by pedestrian detection tasks.The main contents are as following(1)For the multi-scale problem in pedestrian detection,a multi-scale pedestrian detection algorithm based on feature fusion is designed in this paper.First,based on the Faster RCNN model,a multi-scale feature fusion network combined with residual feature enhancement is designed to enable the feature map to have both high semantic and rich detailed texture information,and to reduce the information loss phenomenon caused by channel down scaling in the fusion process.Secondly,the loss function in the algorithm is modified by adding the KL loss branch to the localization loss branch to establish probability distributions about the prediction boxes and the real boxes,respectively,and calculating the KL scatter between them to shorten the distance between the prediction boxes and the real boxes.Finally,the two parts of the improved method are fused to perform the detection of multi-scale pedestrians.(2)For the occlusion problem in pedestrian detection,this paper designs an attention mechanism-based algorithm for occluded pedestrian detection.First,based on the Faster RCNN model,a channel attention mechanism combining global average pooling and maximum pooling is designed to guide the network so that the vectors generated by the feature map after the squeezing and excitati on operations correspond to the channels,and guide the network to focus on the visible part of the pedestrians in the image as a way to reduce the background occlusion of pedestrians.Finally,the two parts of the improved method are fused to perform dete ction of obscured pedestrians.(3)The feature fusion-based multi-scale pedestrian detection algorithm and the attention mechanism-based occlusion pedestrian detection algorithm are experimentally validated on the Caltech dataset and the City Person dataset,respectively,and the results are compared with existing pedestrian detection algorithms.The results show that the feature fusion-based multi-scale pedestrian detection algorithm reduces 1.4% and 1.7% of missed detections compared with the current optimal algorithm on the Caltech dataset and City Person dataset,respectively;the attention-based occlusion pedestrian detection algorithm reduces 1.8% and 0.3%of missed detections compared with the current optimal algorithm on the City Person dataset and Caltech dataset,respectively.This shows the effectiveness of the proposed method. |
