Video Surveillance-Based Detection Of Urban Road Traffic Accident

Video-based urban road traffic accident detection is an important task of smart transportation.Studying an accurate and timely automatic detection algorithm for traffic accidents based on computer vision technology is of great significance.In the case where most high-quality traffic video datasets are classified,to solve the difficulty of deep learning model training caused by data missing,this dissertation proposes a highly scalable production framework based on the CARLA traffic simulation simulator,which can produce high-quality traffic accident simulation data,and creates a large-scale videobased traffic accident monitoring simulation dataset.This dataset provides data support for traffic accident detection.Additionally,when applying conventional video anomaly detection methods to solve traffic accident detection problems,it is difficult to distinguish between normal and abnormal traffic behavior characteristics in traffic videos.This dissertation proposes an attention mechanism algorithm based on video saliency guidance,which provides more robust video action aggregation features for multipleinstance learning algorithms to enhance the detection capability of traffic anomalies.Finally,regarding the difficulty in distinguishing non-accidental and accidental behaviors in traffic anomalies at a fine-grained level,this dissertation proposes a traffic accident search algorithm based on future frame reconstruction method,and combines the simulation dataset and attention mechanism based on video saliency guidance proposed in this paper to optimize the traffic accident detection task.The main contributions of this dissertation can be summarized as follows:(1)This dissertation proposes a highly scalable dataset production framework based on the traffic simulation tool CARLA and implements the preparation of a largescale dataset.With the support of the CARLA simulation tool API,this framework includes the design of higher-order traffic environment background setting components and traffic control components,which enable the simulator to randomly generate traffic accidents.Based on this,a large-scale urban traffic video dataset of more than 20 hours,including both normal and accident videos,is generated.In addition,this dissertation conducts digital image analysis and quality analysis on these videos,demonstrating the ability of the simulation dataset to be transferred to real-world datasets.(2)This dissertation proposes a multiple-instance learning approach that addresses the characteristic differences between abnormal and normal traffic behaviors in traffic videos.Specifically,a video saliency-guided attention model is designed to capture the key features of abnormal traffic behaviors.In addition,a regularization method based on dynamic time warping algorithm is developed to improve the stability and robustness of the model during training,leading to successful detection of abnormal traffic behaviors.This approach resolves the challenge of applying weakly-supervised video detection algorithms to traffic scenes,providing an effective solution to the problem of traffic anomaly detection.(3)This dissertation proposes a traffic accident search algorithm based on encodingerror reconstruction.The algorithm compresses the sample space and performs deeper classification and detection of whether traffic abnormal video segments are accidents at a more finer granularity level,addressing the problem of the weaklysupervised accident detection algorithm proposed in(2)being unable to determine whether traffic abnormal behaviors are accidents or not.