Road Hazard Vehicle Detection Based On Deep Learning

Road dangerous vehicle violation detection has a wide range of application value in the field of intelligent transportation.At present,researchers mostly use convolutional neural network(CNN)to detect dangerous vehicles on the road.However,the existing detection framework still has the problems of substandard detection accuracy and slow inference speed.The Transformer model is a new architecture based on attention mechanism,which has the advantages of paying more attention to global information and modeling long-distance dependence ability than CNN,which has brought major innovation to the field of object detection.With this in mind,this article combines Transformer with existing inspection frameworks to balance inspection speed and accuracy.Study the rapid and accurate detection of dangerous vehicles on the road,the main contents are as follows:Aiming at the fact that the existing single-stage object detection algorithm YOLOX detector has fast inference speed but still lacks detection accuracy,the YOLOX model with attention mechanism is studied.Firstly,the YOLOX backbone network is replaced with Conv Ne XT to improve the feature extraction ability of the model.Then,the hybrid domain attention module is added in the feature fusion stage,and the vehicle information is paid more effectively in the spatial and channel dimensions of the feature map.Finally,it is compared and analyzed with the existing detection algorithm.Experiments show that the detection accuracy of the improved YOLOX model reaches 0.967,and its performance exceeds that of other commonly used detectors.At the same time,compared with the ablation experiment of YOLOX algorithm,the average accuracy of the YOLOX model established in this thesis is improved by 3%,which verifies the effect of Conv Ne XT network and mixed domain attention module.It is aimed at the problem that the detector of the existing two-stage object detection algorithm has high detection accuracy but slow inference speed.Using the classical two-stage detector Faster RCNN as the baseline model,the fusion Transformer and self-supervised pre-trained object detection framework are studied.Firstly,the backbone network of Faster RCNN is replaced with a feature extraction network based on Transformer design to improve the global information attention ability of the model.Then,aiming at the poor performance of visual transformer on small-scale datasets,slow convergence speed and easy overfitting of models,self-supervised pre-training based on mask learning is used to guide the network to learn vehicle feature representation.Experiments show that the detection accuracy of the Faster RCNN model based on mask learning reaches 0.980,and its performance exceeds that of other commonly used detectors.At the same time,compared with the ablation experiment of Faster RCNN’s algorithm,the average accuracy of the model established in this thesis is increased by 5%,and the calculation amount is reduced by 10% compared with the basic model,which verifies the effect of Transformer architecture and self-supervised pre-training.The research method is packaged and packaged,and the road dangerous vehicle detection software is developed,which can realize the visual management and real-time detection of dangerous vehicle data.The road hazard vehicle detection system developed based on the Py Qt5 platform has a user-friendly interface and simple operation,which promotes the application of artificial intelligence technology in the field of traffic management.The thesis has a total of 43 figures,9 tables,and 87 references.