Cross-view-oriented Traffic Light Detection

With the development of deep learning,autonomous driving technology has received increasing attention from industry and academia.The current autonomous driving technology mainly relies on lidar for target perception.This method is not only costly,but also difficult to discriminate such semantic targets as traffic lights and traffic signs.Therefore,adopting the method based on computer vision with the help of cameras is of great significance to solve the problem of detecting traffic lights under different viewing angles.This article focuses on the detection and perception of traffic lights in real street scenes,the detection of extremely small targets encountered in cross-view traffic light detection,the problem of domain adaptation between traffic light in different domains,and the migration between complex scenes.Solutions to the three problems are given and the main research results obtained include:(1)A two-way regression traffic signal detection method based on attention mechanism is proposed.By introducing the attention module in the network,the weight of the region of interest in the feature map is increased.In addition,a two-branch regression network is designed to detect small traffic light,which combines the detection results of traffic lights and traffic light frames.Experiments on the VIVA traffic signal data set show that the m AP of the proposed method can reach 51.14%,which is significantly better than the existing general target detection methods.(2)A domain adaptive target detection method for cross-view traffic lights is proposed.By introducing the domain adaptation module on the basis of the original target detection framework,the semaphore detection model can be adapted to the data distribution of the target domain by means of confrontation training.Thereby,the problem of reduced detection effect caused by different cross-domain data distribution under different perspectives is solved.Without using any additional target domain annotation data,a more robust target detector can be trained for the target domain.Experiments conducted on traffic light datasets under two different scenarios of VIVA and LARA show that the performance of the proposed method is significantly better than the general target detection method.(3)A light-weight target detection method based on the generalized merging ratio(Generalized Intersection over Union,GIOU)loss function is proposed.The feature extraction network is compressed by the channel pruning strategy,and the GIOU loss function is introduced in the training phase,and the smooth non-maximum suppression(Soft Non Maximum Suppression,Soft NMS)is introduced in the post-processing phase.Therefore,the accuracy of the network is almost lossless while the calculation amount is greatly reduced.It solves the problems in the actual application that the network calculation is large and the training model is difficult to deploy on the embedded side.Experiments on a private pedestrian vehicle data set show that m AP still remains competitive when the network model FLOPs is reduced by a factor of three.