Obstacle Recognition System Design Based On Convolutional Neural Network

As a vulnerable group,the visually impaired often encounter difficulties that are unimaginable in ordinary life.At present,the blind mainly rely on the traditional way of guide.However,most blind sidewalk in cities have certain safety hazards due to planning or human factors.The detection range of blind canes is limited,and it is difficult to detect protruding dangerous objects.Guide dogs have long training cycles and high costs.These traditional methods have the disadvantages of low technology content,simple function and limited detection.In order to make the blind walk safely and conveniently,this paper has carried out research on the intelligent guide blind system.Because the traditional machine learning algorithm requires a large number of manual feature extraction,the training is time consuming and laborious,and it is difficult to meet the real-time detection of the target.With the development of deep learning,convolutional neural network has demonstrated its strong learning ability in image classification and other fields.Therefore,this paper designs an obstacle recognition system based on convolutional neural network.The research topic also applies to manless driving and other fields.Considering the guide function of the system,this research subject involves image processing,computer vision,Text-To-Speech(TTS),TCP/IP and other technologies,the binocular distance measurement algorithm is combined with the convolutional neural network,a recognition system with functions of classification and ranging is designed.The main work content is as follows:This paper first introduces the principle of the traditional target detection algorithm and the target detection algorithm based on convolutional neural network are introduced in detail,the shortcomings of traditional target detection algorithm are analyzed,and several mainstream detection algorithms based on convolutional neural network are compared and analyzed experimentally.The YOLOv3 algorithm is selected as the basic network of obstacle recognition system.In order to solve the problem that YOLOv3 missed detection and false detection in the night detection task,a data augmentation layer based on MSRCR(Multi-Scale Retinex with Color Restoration)algorithm is added after the network data reading layer.The network is pre-trained on the mixed dataset of VOC2007 and VOC2012,after obtaining the initialization weight,finetuning it on the obstacle dataset.The multi-scale training is used to obtain the recognition model,and a comparison experiment is conducted with several mainstream target detection algorithms on the self-made obstacle data set.The experimental results showed that the mAP of the algorithm in the night detection task is 7.4% higher than YOLOv3.Secondly,it introduces the overall scheme design of the system,and realizes the obstacle identification system with the embedded system and cloud server architecture,which can guarantee the real-time performance and reliability of the system.The system design mainly includes two parts: embedded system and cloud server.The embedded system uses Raspberry Pi 3 Model B as the main controller,and the peripheral modules mounted on it include: image acquisition and voice broadcasting module.The image captured by the front end interacts with the cloud server through socket communication.After being processed by the algorithm on the cloud server,the result is sent back to the front end in the form of text,which is processed by the speech synthesis chip and then broadcast to remind the blind to avoid.The cloud server adopts high-performance GPU and CPU.It is mainly used to process the images collected by the front end and obtain target classification and ranging results.In addition,this paper introduces the principle of the stereo matching algorithm in detail,and completes the calibration and correction experiments for the binocular camera.Finally,the feasibility and real-time performance of the obstacle recognition system are verified through practical tests.