Research And Implementation Of Road Traffic Anomaly Detection System Based On Deep Learning

With the rapid development of the transportation industry,the traffic flow in our country is increasing,but at the same time,there are also various abnormal behaviors in the road traffic process.At present,with the help of the powerful road traffic monitoring system,it is relatively popular to use technologies of computer vision for anomaly detection.However,many algorithms only work on the pixel dimension and train models in an unsupervised manner,which means their information support is relatively single,making it difficult to distinguish anomalies based on road traffic rules.At the same time,the amount of data generated by video surveillance is relatively large,resulting in high resource requirements for the training process.In response to the above issues,according to the basic principle of video anomaly detection and the characteristics of the road traffic environment,the thesis proposes a road traffic anomaly detection method based on Object-meta,which decomposes the detection task into four dimensions,which are type,location,optical flow,and pixel.Meanwhile,in order to perform differential detection,the thesis proposes a road traffic anomaly detection network based on Multi-level Memory Guided Autoencoder.Finally,a road traffic anomaly detection system was designed and implemented.The main works of the thesis are as follows:1.A road traffic anomaly detection method based on Object-meta has been proposed.Every Object-meta is generated through road segmentation,optical flow calculation,instance segmentation,and feature fusion.Then they will be used as inputs of the anomaly detection network.In this way,the detection process can be supported by multi-dimensional information.Besides,Object-meta remove the irrelevant background areas,thus reducing resource usage during training and reducing the impact of environmental noise.Comparative experiments on multiple networks and datasets show that this method can effectively improve the accuracy of anomaly detection and has good performance.2.A road traffic anomaly detection network based on Multi-level Memory Guided Autoencoder is proposed.On the basis of Object-meta,the thesis constructs the Multi-level Memory Guided Autoencoder.It has multi-level memory modules and uses data-cutting strategies,so as to construct multi-level memory pools.Meanwhile,the group convolution method was used to reduce parameter and computational complexity.The experimental results show that the network can perform differential detection,improve detection accuracy,reduce the model size,and identify the specific dimensions of anomalies,thereby distinguishing the types of anomalies.3.A road traffic anomaly detection system is designed and implemented.The system integrates the Object-meta generation network and the road traffic anomaly detection network,and develops a graphical user interface based on Py Qt framework,so as to provide users with functions of data annotation,Object-meta generation,model training,and anomaly detection.The test results indicate that the system responds promptly and can effectively detect road traffic anomalies,meeting the design requirements.