Research On Sidewalk Scene Recognition Algorithm Based On Deep Learning

People who are blind or visually impaired due to vision defects will encounter various difficulties in their daily lives.At present,the vast majority of visually impaired people mainly rely on traditional blind guides to travel,and the blind channels that sighted people rely on have great safety hazards due to irregular construction or damage and occupation by human factors.The disadvantages of traditional blind guide methods such as single function,small detection range,low technology content,etc.make it difficult for the visually impaired to feed back the information of multiple obstacles on the sidewalk or blind path.In order to travel safely and conveniently for the visually impaired,this paper conducts research on the identification of common obstacles on the sidewalk,and designs an effective sidewalk scene obstacle recognition algorithm to assist the visually impaired to travel.The main tasks completed in this paper are:1.Image preprocessing and image denoising.First,it briefly explains the theoretical basis of deep learning and image processing,as well as the composition and working principle of convolutional neural networks.Then the data set image is preprocessed,and the noise in the image is removed by a method based on the combination of wavelet transform and median filter.2.Propose a target detection algorithm optimization model based on Faster RCNN.On the basis of Faster RCNN,three improvements have been made.The first is to optimize Res Net to reduce training parameters,and to improve the speed of network training and detection;secondly,to improve the RPN structure,by adding a3*3 convolution,increasing the receptive field of the network and increasing the number of Detection accuracy of scale targets;Finally,replace NMS with Soft-NMS to avoid directly deleting anchor frames that exceed the threshold and reduce the missed detection rate of the model.Through comparative experiments on multiple models,it can be seen that the improved Faster RCNN-V is suitable for the detection of sidewalk obstacles and enhances the robustness of the model.3.An improved YOLOV3-N target detection algorithm model is proposed.Use YOLOV3 as the basic network,for its optimization,and replace traditional convolution with deep separable convolution in the residual structure,which greatly reduces the parameters of the structure and improves the detection accuracy and speed of the YOLOV3 algorithm for sidewalk obstacles;Then,when selecting the anchor frame,the Soft-NMS algorithm is used instead of the NMS algorithm,and DIo U is introduced.By combining the two,it is judged whether the two targets in the higher overlap part are the same target,so as to effectively suppress the frame and improve the detection accuracy of dense obstacle targets in sidewalk scenes.The results show that,compared with the algorithms SSD,YOLOV3,YOLOV4,Faster RCNN,the improved YOLOV3-N model has better obstacle target detection effect.