Research On Pedestrian And Vehicle Detection Based On Anchor-free

Object detection task is a very important task in the field of computer vision,which is the foundation of other computer vision tasks.The research of object detection methods has been greatly improved from the traditional machine learning method to the method based on deep learning.At present,almost all the most advanced object detection methods based on deep learning rely on the predefined anchor.However,there are some problems in the Anchor-based object detection methods,such as manually designed anchor,the imbalance between positive and negative samples,and so on.Therefore,this paper abandons the predefined anchor and uses Anchor-free object detection for pedestrian and vehicle detection in security field.The Anchorfree object detection method uses the pixels in the center region of the object as training samples to predict the offset of the bounding box pixel by pixel.However,the current Anchor-free object detection method has the problem of low accuracy.Firstly,in order to solve the problem of low accuracy in the Anchor-free object detection method and get a real-time and efficient network,this paper redesigns the network,including the backbone network,neck layer,detector head,sample selection method,learning rate and post-processing.In this paper,the backbone network adopts the method of adding Res2 Net module to the Res Net to increase the receptive field,the neck layer adopts the path enhanced feature pyramid PAFPN in the PANet which is in instance segment field to enhance the feature,the detector head adopts GFocal Head V2.The DFL loss function is used to solve the boundary uncertainty.The QFL loss function is used to solve the inconsistency between classification score and quality score of bounding boxes in training and forward.The sample selection method uses ATSS to solve the imbalance between positive and negative samples.Soft-NMS is used in post-processing to solve the problem of adjacent bounding boxes being suppressed.In this paper,the algorithm of each module is completely described,and the related experiments and analysis are carried out.The results show that the network designed in this paper is efficient.Secondly,in the stage of object classification,there is a mismatch phenomenon of different classification results for the same object in the scene.In this paper,according to the similarity characteristics of different classification objects,an optimization algorithm is proposed,which is based on the Io U elimination strategy,called similar object suppression rule.In the postprocessing stage,according to the degree of Io U overlap of the two bounding boxes and the similarity characteristics of different classification objects,the rule algorithm can judge whether the bounding box is retained or not,and eliminate the bounding box with low confidence.The improved algorithm is verified on the proposed pedestrian and vehicle dataset.The results show that the similar object suppression rule can get more accurate detection effect and reduce the false positive rate.Finally,the network designed in this paper and the proposed optimization algorithm are based on the pedestrian and vehicle dataset of real security scene for training and test evaluation.Based on this dataset,the experimental results show that the Mean Average Precision(m AP)of the improved algorithm is 0.704,which is 0.031 higher than that of the original scheme.Moreover,in terms of speed,it can achieve 30 FPS,and achieve a high precision real-time effect in pedestrian and vehicle scenes based on Anchor-free object detection method.Compared with the Anchor-based object detection method,the network is simple and flexible,and can adaptively predict the bounding box of pedestrian and vehicle.The proposed algorithm provides high precision and low false positive rate for general target detection algorithms.The algorithm is effective and paves the way to the development of pedestrian and vehicle detection algorithm in real-time.