Research On Pedestrian Detection Based On Fused Shallow Feature And Convolutional Neural Network

In recent years,pedestrian detection system has made breakthrough progress due to the powerful representation ability of convolutional neural network.Generally,the shallow feature map of convolution neural network represents details and shapes information of the image,while the deep feature map represents the semantic information of the image.Moreover,both the deep feature and the shallow feature are indispensable components in the task of pedestrian classification and location.However,most of the existing pedestrian detection methods focus on fusing shallow features and deep features.Which not to further analyze the fused feature and result in irrelevant information of the extracted pedestrian features.It also makes the false detection rate not reach the lowest level.To solve this problem,this thesis starts with the improvement of the existing classical convolution neural network,uses the related techniques to extract the shallow and deep features of the pedestrian,and processes the post-fusion characteristic channel by a certain means,and finally achieves the purpose of improving the performance of the pedestrian detector.The main work of this thesis is as follows:Using the Region Proposal Network(RPN)and Fast RCNN(FRCNN)as the overall detection framework,a pedestrian detection method based on adaptive feature channels are proposed.The improvement of this method mainly focused on the deep feature map of the target.The main improvements are as follows:(1)Thesis designed SFCM module is used to extract the shallow features of the convolution neural network and to integrate them into the deep features,improving the comprehensiveness of the features represented by the deep feature channel.(2)In the deep feature channel after fusion,the AFCM module is applied to update the corresponding feature map value through the extrusion,excitation and weighting calibration function of the module.In order to realize the pedestrian detection method of adaptive feature channel,the network can effectively learn strong discernment pedestrian features to classify and locate tasks.In order to verify the effectiveness and superiority of this improvement point,this thesis evaluates the effectiveness by adding the proposed modules one by one,and compares the performance of the proposed method with that of the existing pedestrian detection methods on the open data set.The experiment shows that this method has certain advantages.In multi-task learning,auxiliary tasks can promote better learning features of the model.Therefore,in order to further optimize the detection model,this thesis propose a pedestrian detection method combining weak classification supervision and adaptive feature channel based on the above detection algorithm and combined with multi-task learning technology.The improvement of this method still focuses on the deep feature mapping,and its main improvements are:(1)The Body module is optimized by replacing the pooling operation and the convolution operation with the dilated convolution.Therefore,the network can effectively avoid the loss of small-scale pedestrian information caused by the pooling operation and enlarge the receptive field of the image.Consequently,the performance of small-scale pedestrian detection can be improved.(2)Taking the area of pedestrian detection box as the supervision information,a weak classification supervision branch was added in the AFCM module,so as to enhance the model's ability to learn pedestrian characteristics.(3)This method realizes end-to-end multi-task training.The experimental results show that the proposed algorithm is competitive.