In order to solve the problems of poor detection effect of some abnormal vehicles in night scenes and large amount of model parameters, a night vehicle abnormal dataset was made and YOLOv5 was improved. The abnormal images of vehicles in the night were put into the training, and the final detection effect was improved. The improvements include: in the Backbone layer of YOLOv5 network model, the extended convolution depthwise over-parameterized convolutional with deep overparameterized convolution was introduced, which could accelerate the convergence speed, improve the network performance, effectively reduce the amount of parameters and floating point calculation and improve the detection accuracy. according to the lightweight convolution GSConv, which could achieve a balance between the accuracy and speed of detection, the new VOV_SimAM module and C3_SimAM module were designed to improve the extraction ability of important abnormal features. The flow alignment feature pyramid networks cross-fusion pyramid structure and FaConcat splicing method was proposed to reduce the amount of parameters and floating point calculation. The Wise-Intersection over Union Wise-IoU loss function was introduced to solve the balance problem of good and poor quality samples, and the detection of vehicle collision, vehicle fire and vehicle overturning was realized. Compared with the original YOLOv5 network model, the amount of parameters and floating point calculations are reduced by 8.5% and 28.8%, respectively. The detection average precision AP values of vehicle collision, vehicle fire and vehicle overturning are increased by 11.3%, 5.9% and 29.1%, respectively, and the mean average precision mAP value is increased by 15.4%, providing a good reference for intelligent video policing applications.