UDBNet: Detecting Unsafe Driving Behaviors Relevant to Urban Traffic Safety

Recently, the number of traffic accident deaths caused by unsafe driving has been increasing dramatically. At present, unsafe driving behavior detection in urban traffic remains a challenging task due to complex background, low spatial resolution, and poor illumination conditions etc., which limit generalization ability and low detection accuracy. In this paper, a novel detection network, called UDBNet, is proposed for unsafe driving behavior detection in complex traffic monitoring scenes. UDBNet introduces a deformable global convolution block to overcome the shortcomings of the existing Convolutional Neural Network (CNN) in capturing global features and adapting to deformed targets. In addition, UDBNet adopts a new Feature Pyramid Network (FPN), termed MBFPN, to enhance the ability of feature fusion. Finally, the effectiveness, robustness, and generalization of UDBNet are proved through extensive experiments.