Image Detection Based On Vehicle Embedded System

Image object detection,as one of the basic problems in computer vision,has made great progress after the addition of deep learning.Object detection algorithm based on convolutional neural network(CNN)has been successfully applied in industrial detection.In the field of driving,visual images can provide rich information for vehicles in the process of driving.Image target detection plays an important role in assisted driving and autonomous driving tasks.However,the on-board embedded target detection system has higher requirements for the real-time and detection accuracy of the target detection,and in the process of auxiliary or automatic driving,the on-board embedded system needs to process a variety of data at the same time to make decisions or provide suggestions on the constantly changing road conditions.Therefore,the vehicle embedded system has the characteristics of high concurrency,and how to take into account the detection accuracy and detection speed is the research focus in the industrial application field.Based on deep learning,this paper carries out research on image target detection of vehicle-mounted embedded systems,with specific contents as follows:Based on the common targets in driving tasks and COCO data set,a data set containing12 common targets,73000 mark data and 39,000 images in driving tasks is constructed through expansion,merging and screening.Hysi 3559 A intelligent chip is used as the test platform,RTX4090 on Linux system is used as the training computing platform,YOLOv4,which is the representative in the field of target detection industrial application,is taken as the benchmark,and the test results of YOLOv4 on Hysi 3559 A are taken as the benchmark,aiming at the real-time operation requirements of the vehicle embedded system.Center Net target detection algorithm without anchor frame was used as the improved experimental algorithm,which was improved from three aspects: model structure,training strategy and model compression.In order to improve the capability of feature extraction,the Center Net backbone feature extraction network was replaced by CSPDark Net53 in the model structure to improve the capability of feature extraction.In the feature fusion stage,the feature pyramid structure was adopted to enhance the capability of feature fusion.In view of the problem of small number of positive samples in model training,In the training strategy,the positive and negative sample differentiation method of transition samples is used,which alleviates the problem that real tags are forcibly mapped to [0,1],which leads to the slow convergence speed during network training.For the feature pyramid structure,Kmeans clustering analysis was used to classify the feature targets of different sizes,and three types of large,medium and small targets were selected and allocated to different feature layers respectively.Elliptic Gaussian kernel function was used to replace Gaussian kernel function for data sampling optimization.Mosaic data was used to enhance the robustness of the model during training.In terms of model compression,the channel pruning method was used to compress the trained model to achieve accuracy,remove redundant weight parameters to improve the calculation speed of the model,and knowledge distillation was used to improve the detection accuracy of the model after pruning.The experimental results on Hysi 3559 A show that the detection accuracy and speed of the improved Center Net are better than that of the YOLOv4 target detection algorithm.The m AP reaches 95.3% at 0.5 confidence,and the compressed algorithm model reaches a real-time running speed of 55 FPS on Hysi 3559A.