Research On Abnormal Vehicle Behavior Detection Based On Machine Vision

Abnormal vehicle behavior refers to the behavior of a vehicle that deviates from the behavior of the normal driving vehicle or the normal movement pattern in a specific road scene.The detection of abnormal vehicle behavior is an important part of traffic monitoring system,and automatic real-time detection of abnormal vehicle behavior in monitoring scenarios is an important development direction in the field of intelligent transportation.Abnormally behaving vehicles may cause serious traffic accidents,and traffic accidents will inevitably lead to abnormal vehicle behaviors.Therefore,the effective detection of the abnormal vehicle behaviors in important traffic places helps to improve the safety and operation efficiency of road traffic,and helps to reduce the casualties and economic losses caused by traffic accidents.The detection of abnormal vehicle behavior is a complex problem.The detection methods mainly include induction coil detection,radar signal detection,GPS signal detection,machine vision detection,etc.The machine vision-based method is widely utilized because of its low cost,high deployment flexibility,and good information processing effects.When using machine vision to detect abnormal vehicle behavior,we need to detect the vehicles in the video and extract their motion characteristics.Then based on further analysis of the vehicle motion characteristics,a vehicle motion pattern model is constructed.Finally,the abnormal behavior of a vehicle is detected by analyzing the matching relationship between the real-time trajectory of the vehicle and the motion pattern model.Although the research on abnormal vehicle behavior detection has made great progress,there are still many problems to be solved.Firstly,although the real-time object detection models based on deep learning,such as the yolov5 s and yolov5 m,are capable of most object detection tasks,they do not have high recognition accuracy for these vehicles that are small and partially occluded or truncated.The reason is that the feature extraction ability of the models is insufficient.Secondly,in the actual scenes,because of the vehicle trajectories with different dimensions,the complex trajectory patterns,and the huge trajectory data,it is difficult for the existing clustering methods and probabilistic statistical learning methods to obtain good trajectory pattern learning results,and the results conform to the physical structure of the road.Thirdly,due to the lack of a reasonable method to unify the trajectory dimensions,it is difficult for vehicle behavior analysis methods based on motion trajectory to make use of the key spatio-temporal characteristic information in the original trajectory.Which leads to low detection accuracy and insufficient recall rate of abnormal trajectory,and it is difficult to detect abnormal trajectory across multiple motion modes.Finally,most of the work related to the detection of vehicle behavior by traffic management system relies on manual work,and there is still insufficient research on the intelligent real-time detection of abnormal vehicle behavior.This article has launched researches on the aforementioned problems,and its contents are as follows.(1)The problem of different stages and multi-scale feature fusion based on parallel convolutional blocks is studied.Convolutional neural network is the backbone of vehicle detection model based on deep learning,and its performance is determined by its feature extraction ability.Classification accuracy is an important measure of feature extraction ability of convolutional neural network.Hence,classification accuracy is an important measure of vehicle detection model performance.Aiming at the problems that increasing the network depth to improve its classification performance will lead to a huge increase in the amount of model parameters and model over-fitting,this study improves its classification performance from the perspective of increasing the width of the convolutional neural network.This study first adds a parallel auxiliary convolution block with different sizes in the previous stage of the network to extract low-level features with different sizes,and then fuse these features with the features of different levels of the network.And a strategy which fuses multi-scale features in the previous stage of the convolutional neural network for improving its classification performance is proposed.Experimental results show that this strategy can significantly improve the classification performance of existing high-performance convolutional neural networks.(2)A real-time object detection model specially used for vehicle detection is proposed.When the general object detection model detects vehicles in the image,the detection accuracy of the small-sized vehicles that are partially occluded or truncated is not high.From the perspectives of improving the model’s feature extraction ability and of improving the model’s detection accuracy for the small-sized vehicles,this research designed a real-time detection model specifically for vehicle detection.First,the model exploits the feature fusion strategy based on the parallel convolutional blocks,the cross-stage partial connections structure,the spatial pyramid pooling structure and the feature pyramid network structure to enhance the feature extraction ability of convolutional neural networks.Then,a dedicated small-size vehicle detection layer is designed to detect the small-size vehicles.This layer uses the super-size feature map in the network to generate priori boxes,and only generates a small-size priori box on each pixel of the feature map.At the same time,this layer uses a smaller number of convolution kernels to achieve classification and regression.Experimental results show that the model has the advantages of high detection accuracy,fast reasoning speed,and small model size.(3)A learning framework for normal vehicle trajectory pattern is proposed.The vehicle trajectory dataset obtained in real monitoring scenarios have the characteristics of large number and huge dimension difference.Due to the lack of effective trajectory processing methods,it is difficult for existing normal trajectory pattern learning methods to utilize the key spatio-temporal feature information contained in trajectory samples.In addition,it remains to be solved how to determine the reasonable value of the hyperparameter of the Density-Based Spatial Clustering of AP-Plications with Noise(DBSCAN)algorithm when the similarity between trajectory samples is very different.To solve these two problems,this study proposed a vehicle trajectory dimension reduction algorithm and an algorithm for obtaining the initial values of the hyperparameters of the DBSCAN algorithm,and then proposed a normal vehicle trajectory pattern learning framework.Experimental results show that the trajectory clustering results obtained by this framework are well and conform to the constraints of road physical structure and traffic laws and regulations.(4)A dual-model parallel detection algorithm for abnormal trajectories is proposed.The existing models are difficult to identify the category of the incomplete vehicle trajectories,and cannot effectively detect the abnormal trajectories across multiple motion patterns.On the basis of unifying the dimensions of all trajectories in the trajectory dataset,this research proposes to exploit the high-dimensional hidden Markov model for enhancing the use of the key spatiotemporal feature information of the original trajectories.This method aims to improve the detection accuracy of the model with respect to incomplete trajectories.At the same time,the angle model of the motion direction vector of the trajectory patterns is used to further constrain the detection results of the high-dimensional hidden Markov model.This approach aims to improve the detection accuracy of these abnormal trajectories across multiple motion patterns.Furthermore,a dual-model parallel detection algorithm for abnormal trajectories is proposed.The experimental results show that the algorithm has excellent performance and can significantly improve the detection accuracy of the high-dimensional hidden Markov model.(5)A detection model of abnormal vehicle behavior is proposed.The traffic management system is difficult to intelligently detect the abnormal behavior of vehicles in the monitoring scene timely,and most of the work about abnormal vehicle behavior detection needs to rely on manual completion.To solve this problem,on the basis of reasonably extracting,deleting and saving real-time vehicle trajectory data,this paper proposes an abnormal vehicle behavior detection model based on trajectory analysis.This model combines the dual-model parallel detection algorithm of abnormal trajectory and virtual coil speed measurement method.In this model,overspeed detection and abnormal trajectory detection are performed simultaneously after the vehicle trajectory is extracted.The experimental results show that the model can automatically detect abnormal vehicle behaviors such as illegal lane changing,reversing,occupying non-motorized vehicle lanes,speeding and other complex behaviors in a timely manner.