Research On Urban Traffic Abnormal Behavior Recognition Based On Multiple Instance Learning

With the rapid development of urbanization in China,the number of cars in cities is increasing year by year,leading to increased traffic congestion on urban roads.The accident rate is increasing,and the problem of urban traffic safety is becoming more and more prominent.The core of improving the urban traffic environment is to build a management system that can effectively deal with occasional traffic accidents.Abnormal behavior recognition is the first step of the management process,which is also its foundation and core.However,the traditional recognition method based on the ground sensor coil is simple in function and quite inconvenient in use and maintenance.The road traffic abnormal behavior recognition technology based on monitoring video can make up for the shortcomings of the traditional recognition method,but also can make the road monitoring device play a greater role.Traffic abnormal behavior recognition is a dynamic image processing process,visual modeling and calculation process is more complex compared to static images.Moreover,since vehicle recognition can be disturbed by many factors such as weather environment,designing computationally efficient and robust models is of great significance in practical applications.Therefore,in this paper,the video is used as a package according to the idea of multiple instance learning,and the video is split into video clips as instances.The image and temporal information of the instances are extracted and integrated by visual and temporal feature extraction networks.Finally,the abnormal behavior scores of the instances are predicted using a multilayer perceptron so that the model can recognize traffic abnormal behavior.The main work of this paper is as follows:(1)The traditional ranking loss function cannot rank instances that are not accurately labeled,so this paper designs a multiple instance ranking loss function.This function only needs to know whether the video contains abnormal behavior or not,and does not need to know the specific time when the abnormal behavior occurs and ends to iterate and optimize the model.In order to get a better recognition effect,this paper also adds sparsity constraint and smoothness constraint in the model according to the characteristic of the small probability of occurrence of traffic abnormal behavior,so that the model can locate the time of occurrence and end of abnormal behavior more accurately.(2)The traditional convolutional neural network is prone to the problem of gradient disappearance and can only extract the spatial features in the image.Therefore,this paper uses the residual network as the backbone network of the visual feature extraction network.In order to better extract the spatiotemporal features in the video,this paper uses 3D convolution instead of 2D convolution.And we use the densely connected network and soft thresholding to reduce the number of parameters and computation of the model and improve the efficiency of the network to extract features.(3)The long short-term memory network can effectively extract the temporal features in the video and weaken the influence of redundant information on recognition results by attention allocation for the redundant information in the sequence.However,it cannot solve the back-and-forth dependency problem in the sequence.Therefore,in this paper,we analyze the algorithm and principle of the long short-term memory network and improve it into a bidirectional long short-term memory network.So that the model can better extract the before and after temporal features of the video after extracting the visual features,and improve the recognition accuracy of the model.(4)In this paper,an urban traffic abnormal behavior recognition system is designed based on a multiple instance learning model.This system can recognize the specific time period when abnormal traffic behavior occurs in the video based on the input video content,which verifies the feasibility of the model designed in this paper.