Research On Pedestrian Detection Algorithm Based On Anchor-free

With the increase of the number of cameras on the street and the demand for automatic driving in our country,the performance requirements of pedestrian detection algorithm are constantly improving.After years of research,pedestrian detection algorithm is becoming more and more mature and can be used in the practical production.After studying the current mainstream pedestrian detection algorithms,we summarize the existing detection algorithms.After comparing different algorithms,we propose a one-stage detection algorithm based on anchor-free,which makes the detection model lighter and suitable for occluded scenes while ensuring certain detection accuracy.In this paper,we improved the general detection model on the anchor and loss function.The lightweight feature extraction network is used to extract different feature layers for fusion.The anchor free technology is used to remove the limitation of the range of the anchor detector in the prediction box,the cumbersome parameter setting and the increased amount of calculation.The final feature layer retains the information of different depth feature layers at the same time,and the algorithm optimizes the labels of positive samples and negative samples by dividing regions,which improves the accuracy of model training.In addition,by analyzing the effectiveness of different detection frames,the visual part of the dataset annotation is introduced and added to the model training as an additional loss to reduce the interference of the background on the detection algorithm training and alleviate the influence of the occluded objects to a certain extent.Finally,the proposed algorithm achieves the reduction of model parameters and the improvement of model speed to a certain extent,and the detection accuracy is also competitive.Aiming at the occlusion problem in pedestrian detection scene,the detection algorithm is improved according to two situations: the pedestrian which occluded by objects or between pedestrians.The rejection loss is added to the loss function of the algorithm,and the IOU is improved to make the prediction frame as far away from other real target frames as possible during the training process,so that the detection frame can avoid the influence of other pedestrians.In addition,according to the characteristics of pedestrian and pedestrian detection,in the process of model training,the real target frame is weighted by blocks,which makes the detection algorithm pay more attention to the learning of the easier part of the head and body,reduces the influence of occlusion,and solves the problem of pedestrian being occluded by objects.Finally,the experimental comparison and algorithm performance comparison are carried out on the universal pedestrian data sets Cityperson and Caltech.Through the comparison,it is proved that the method has certain competitiveness in improving the detection effect and reducing the occlusion problem,and can adapt to more complex pedestrian scenes.