Research On Lightweight And Mobile Deployment Method Of Road Target Detection Algorithm Based On Deep Learning

Road target detection is a key issue and research hotspot in the field of intelligent transportation and automatic driving.In recent years,with the rapid rise of deep learning technology,the road target detection algorithm using convolutional neural network has greatly improved in accuracy and processing time compared with the traditional detection algorithm using sliding window or image segmentation technology,and there are still some problems.Although the complex convolution neural network algorithm design can improve the accuracy of road target detection,the complex algorithm design brings excess calculation and parameter quantity,as well as too large model volume,which leads to higher standard requirements for the computational power of computing equipment.However,in the current era of mobile Internet,the algorithm is difficult to deploy and apply directly on mobile devices and embedded devices with limited computing resources.At the same time,the detection and reasoning speed of the algorithm on mobile devices with limited computing power is poor,resulting in limited application value.Therefore,this thesis carries out in-depth research,improves the complex network structure design of road target detection algorithm based on deep learning,reduces the requirements for computing resources,maintains the detection effect,improves the detection speed,and deploys and applies it on mobile devices.Among them,the specific work contents of this thesis are as follows:1.After studying the one-stage target detection algorithm with detection accuracy and real-time detection effect,based on the yolov5s target detection algorithm and the implementation principle of mobilenetv2 lightweight network,this thesis introduces the residual structure of deep separable convolution and inversion,and carries out lightweight optimization in the feature extraction layer to solve the problems of large amount of parameters and information loss,Reduce the size and parameters of the algorithm while maintaining the detection effect.The map indexes of the improved algorithm in this thesis on COCO and VOC data sets reach 30.1% and 70.1% respectively,which has good performance compared with other lightweight target detection algorithms.At the same time,the detection speed comparison experiment on mobile devices shows that the detection speed of this algorithm is better than other lightweight target detection algorithms.2.The deployment and application of the road target detection algorithm based on deep learning in the mobile terminal are designed and implemented.The road target detection is completed by calling the camera in the mobile terminal and identifying the photos in the gallery,which can meet the needs of real-time detection.