Crowded Pedestrian Detection Algorithm Research Based On Deep Convolutional Network

Pedestrian detection is an important research direction in the field of computer vision.It is widely used in intelligent security,intelligent transportation,pilotless automobile and other fields.However,in the real scene,the background is complex including the dense pedestrians,mutual shelter and lopsided scale,which restricts the detection accuracy of SSD,Faster R-CNN and other target detection method based on deep learning.Therefore,this paper will mainly improve from three aspects:enhancement of depth features,improvement of loss function and integration of multi-scale features,which is of great significance to improve the accuracy and environmental suitability of pedestrian detection.The specific research contents of this paper can be summarized as follows:First,we design a feature extraction network RANet based on deep feature enhancement.The channel attention and spatial attention levels are used to form a cascaded attention module and embedded in the residual module of Res Net50.With the attention of the two dimensions,channel and spatial,not only the influence of local feature information on global features but also the complexity is enhanced.The background descends the features of the human image,and suppresses the background expression in the feature map.Experiments on the RS self-made datasets and the CUHK public datasets show that the miss rates of our method decreased by 2.6% and 1.9%respectively,compared with Res Net50 backbone.Secondly,introduced into the loss function,the repulsion loss between the predicted boxes and the adjacent target’s boxes suppressed the drift of the prediction frame,making the predicted boxes regression more accurately.With Faster R-CNN(RANet)as the benchmark method,an experiment is carried out.The results indicate that the miss rates on the RS and CUHK datasets have decreased by 2.2% and 2.7% respectively after adding the repulsion loss to the benchmark method,which improves the detection accuracy and robustness of the detection model in dense pedestrian scenarios.Finally,we construct a feature pyramid structure DFPN based on multi-scale feature fusion.By introducing shallow features into the feature pyramid structure,the information path between the lower layers and the topmost feature is shortened,so that the rich detailed information and location information of the underlying features can be used to improve the detection performance of multi-scale targets.Finally,the network structure is integrated with the methods in Chapters 2 and 3 of this paper to form the final pedestrian detection model in this paper.Compared with the current mainstream multi-scale pedestrian detection method SAF R-CNN,the miss rates of DFPN on RS and CUHK datasets are reduced by 9.1% and 13.9% on small scale targets,by 3.1% and0.5% on medium scale targets,and by 0.3% and 0.1% on large scale targets.