Research On Obstacle Detection And Recognition Algorithm Of Blind Path Based On Deep Learning

As an important part of urban road planning,the blind road reflects the economy and civilization of a city.However,due to the difference between urban road planning and urban development,the situation of blind road being arbitrarily occupied is becoming more and more serious,resulting in the lack of safety for blind people to travel.How to improve the reliability of travel for the blind has become a hot topic in current research.In recent years,deep learning has made great progress in obstacle detection.The detection and recognition accuracy of network for obstacle targets has been continuously improved.However,the increasing depth of network will lead to the loss of features of small and medium-sized obstacle targets in images,resulting in low accuracy of network detection.Due to the wide variety of blind obstacles,the deep learning network has some problems such as missed detection and misdetection of small obstacles and bar obstacles when detecting obstacles.Aiming at the above problems,based on the summary and analysis of certain research work in the field of blind obstacle detection,this paper realized a YOLO algorithm widely used in the field of target detection,and carried out research from the following aspects:(1)according to the characteristics of the obstacles in this paper,the analysis found small obstacle detection is a difficult point in obstacle detection,use YOLO V3 network based network structure for small obstacles of feature extraction ability is insufficient,and for small offset value is bigger,the obstacle detection frame’s confidence of the box to choose location not accurate enough,lead to the final accuracy is decreased.To solve the above problems,this paper proposes an improved blind track detection method for YOLO V3.First,this paper uses void convolution to replace the 3×3 convolution of the residual structure and the main structure,to increase the receptive field of the convolution kernel,and to expand the extraction range and ability of the convolution check features.Secondly,this paper uses the mixed attention module to focus the Io U region of small and medium-sized obstacles in the feature map in advance,so that the network can extract the features of key points more accurately,reduce the offset value of confidence,and improve the detection effect of small target obstacles.In this paper,the experimental results show that the performance of the proposed network is better than that of the comparison method on the self-built data set.(2)In view of the problem that the backbone feature extraction network has poor feature extraction effect on the bar target,the CSPDark Net53 network structure is taken as the basic model,and the asymmetric convolution structure is integrated into the network to construct a new feature extraction network,so as to achieve effective feature extraction for small targets and bar targets.In order to solve the problem that pyramid pooling module has poor effect on bar and small obstacle pooling in blind path scene,a hybrid pooling module is proposed.In this module,bar pooling has a better pooling effect on bar objects,while square pooling retains the characteristics of square obstacle objects,and bar pooling can further strengthen the pyramidal pooling effect,and further enhance the characteristic information of target cross region.In view of the difficulty of deep neural network in feature extraction to effectively describe the detail features of small target obstacles,and the low accuracy of small target recognition in complex blind scene,a fusion method of multiplying advanced feature information with low-level feature information was proposed to enhance the recognition of small target.In this paper,an improved YOLO V4 network is proposed by integrating asymmetric convolution structure,mixed pooling module and feature fusion method,and the effectiveness of the improved YOLO V4 network is verified on the blind road obstacle data set.Compared with the comparison method,the performance of the proposed method is better.(3)for the blind travel particularity,the design is based on the depth of the raspberry pie learning portable testing device,the trained network model into raspberry pie,camera acquisition obstacles for detection and recognition in image input to the network,determine the type of obstacles,at the same time using the ultrasonic distance measuring current obstacles.Finally,the voice output is used to promote the blind to avoid obstacles reasonably.