Abstract: Aiming at the accuracy requirements of the night vehicle detection model, this paper proposes the night vehicle as the reasearch object and uses convolution neural network in deep learning to construct the detection model. Firstly, the data set is processed with white balance to reduce the interference of street lamp color and enhance the image quality, and mosaic data enhancement is used to enrich the detection data set and improve the detection effect of the model for small target vehicles; Secondly, K-means+〖KG-*3〗+ algorithm is used to select the prior box, and the intersection and union ratio distance is used to cluster the prior box; Then the attention mechanism module is added to the backbone feature extraction network to enhance the channel and spatial feature information of the target in the residual structure feature map; Finally, the gradient equilibrium mechanism is introduced into the original confidence cross entropy loss of the loss function to make the model attenuate the hard and easy samples effectively. Through the experiments and comparative analysis on UA-DETRAC data set, it can be seen that the accuracy of the proposed algorithm can reach 99.24%, and the number of frames per second to process image can reach 19.

Key words: night vehicle detection, deep learning, CBAM, GHM