Research On Multiscale Pedestrian Detection Based On Deep Learning

Pedestrian detection is a key technology in the field of computer vision,involving multiple aspects of knowledge such as image processing and machine learning.As a basic work for other high-level visual tasks such as pedestrian tracking and pedestrian behavior analysis,it is widely used in daily life,providing many conveniences for people,and the research is of great significance.However,facing the issue of pedestrian scale changes,there is still room for improvement.In response to this difficulty,this article conducts research on multiscale pedestrian detection based on deep learning.In terms of datasets,three types of datasets were used to conduct experiments to enrich the diversity of pedestrian detection scenarios and better verify the effectiveness of the model,including the mixed dataset produced from the PASCAL VOC and COCO datasets,the INRIA dataset,and the Caltech pedestrian dataset,and statistical analysis was conducted on pedestrian scales,understand the characteristics of the dataset and the scale distribution of pedestrians.In response to the multiscale problem of pedestrians,the RetinaNet algorithm with feature pyramids and multi-scale prediction was used as the basic model to improve and optimize the network structure for low detection accuracy and high miss rate in multiscale pedestrian detection.In order to improve the ability of feature extraction and obtain rich features,select the multi-branch structure before the feature pyramid to obtain the feature information of different receptive field at different depths.To make the network pay more attention to useful information,attention mechanisms are used to extract important feature information from the network.Through experiments on multiple datasets,it was found that the improved RetinaNet model has improved detection accuracy and achieved real-time detection speed.Moreover,from the detected image results,it can be seen that the improved model can accurately detect pedestrians at various scales in the image,with better results.Aiming at the problem that it is difficult to detect small sized pedestrians in multiscale pedestrians,pre-set anchor frames will limit the detection of pedestrians with variable scales,and have a significant impact on small scale pedestrians.This study was detected by based on the anchor-free FCOS model.In order to improve the detection ability of small-scale pedestrians,a multi-scale feature fusion method is adopted to fuse contextual information of features at different scales.The research results indicate that the proposed method improves the missed and false detections of small-scale pedestrians.Finally,the model was applied in pedestrian videos captured in actual scenes,with high pedestrian accuracy,good generalization ability and practical application effect,achieving the expected goals.