Research And Implementation Of Road Obstacle Detection System In Urban Environment

Automotive assisted driving technology is a pivotal technology at the forefront of the current automotive industry with its inevitable development trend,among which road obstacle detection is one of the technological difficulties.In urban environment,road obstacle detection technology is the safety guarantee of vehicle assisted driving technology with profound research value and significance.Through in-depth research and practical application of road obstacle detection technology,the safety of the driver can be guaranteed to some extent.Based on the target detection algorithm,this thesis designs and implements an effective road obstacle detection system in the urban environment.The main work is as follows:1.Improvements in road obstacle detection algorithm.This thesis employs the YOLO algorithm in road obstacle detection.First of all,after comparing and analyzing the last three versions of the YOLO algorithm,the research chooses the YOLOv5 algorithm.Then,on the basis of the YOLOv5 algorithm,it conducts research from five aspects: pre-training model,algorithm optimizer,data enhancement,loss function and activation function to put forward improvement strategies.By collecting and labeling a total of 9027 pictures,the research forms a data set that conforms to the scene of this system and is applied for training.After plentiful experiments,the research finds that using N-EIo U as the loss function has better improvement for model accuracy.2.Research on optimization of road obstacle detection results.The DeepSort algorithm is used to overcome the shortcomings of large changes in the detection frame of moving objects between different frames when detecting road obstacles in video streams.The ReID model used by the DeepSort algorithm is optimized.The thesis discusses the influence of the algorithm optimizer,loss function,activation function and learning rate scheduler on the ReID model.Based on the experimental results,it concluds that the CELU activation function and the radam optimizer can promote mAP indicators of model respectively.Then resorting the pixel enhancement method to enhance obstacle detection data set,the experiment finds that the accuracy of the YOLOv5 model can be slightly improved.Finally,a road obstacle target determination algorithm is proposed to prompt the driver of the obstacle location.3.A road obstacle detection system in urban environment is designed and implemented.The system can provide users with the functions of picture detection,video detection,video detection tracking,camera detection tracking,and changing storage locations,which exerts decent results in detection accuracy and real-time performance.The system function of this thesis meets the design requirements basically.It can give prompt information of road obstacles in time,thereby effectively preventing the occurrence of traffic accidents.Meanwhile,the system presents an outstanding effect in real-time and performance so that it possesses practical application value.