Traffic Accident Detection For Driving Systems By Adversarial Visual Scene Context Learning

With the rapid development of autonomous driving and advanced driver assistance systems,people have made more and more efforts to improve driving safety when the natural driving state is not clear,especially in the detection of traffic accidents.For example,vehicles and pedestrians that do not obey the traffic lights crossing irregularly,the vehicles in stops suddenly,etc.In the practical driving scene,it is very important for autonomous vehicle to detect the traffic accidents,so as to take measures timely to avoid accidents.According to the problems in traffic accident detection,this paper first proposes a model based on adversarial appearance-action prediction.The novelty of this model is that it can predict the appearance and action characteristics of future frames and future targets from the observed frames,and detect traffic accident by comparing with the RGB pixel-level features of the real frame and the real position of the target.In order to verify the validity of the model,the A3 D data set is used for testing,and the experimental comparison proves that the model can significantly improve the detection effect.Due to the dynamic camera motion and the complex scenes in driving situations,not only the global and local features need to be considered,but also the context of the scene needs to be taken into account to predict the visual relation structure among road participants(pedestrians,vehicles,motorcycles,etc.).On the basis of the appearance-action prediction model,this paper proposes a driving scene accident detection model based on visual context confrontation learning.The model mainly uses a graph-structure embedded generative adversarial network to detect traffic accidents through context conflicts in visual scenes.The network includes a two-branch encoder-decoder module,which predicts the future location of the driving participant and the future frame through the target bounding box and continuous frames of the video sequence.Then the predicted results are combined with the graph convolutional network to learn the relationship between the feature embedding of the object.In order to detect traffic accidents,this paper uses a visual scene context conflict measurement method based on cosine distance.The method proposed in this paper does not require any label annotations and video trimming work,and can be directly used for any original video.Finally,through comparative experiments conducted on the A3 D dataset and the DADA-2000 dataset,the results prove that the model can improve the accuracy of traffic accident detection and is superior to other methods currently proposed.