Nighttime Vehicle Detection Algorithm Enhanced by NightvisionGAN

To solve the problem of low vehicle detection accuracy caused by inconspicuous nighttime vehicle features and complex environmental lighting interference, In this paper, a nighttime vehicle detection algorithm based on NightvisionGAN is proposed. The generator coding and decoding process are connected in corresponding layers, and the residual module is used to realize the conversion of features of different scales, and the CBAM attention mechanism is introduced to make the generator pay attention to vehicle features adaptively, improve the generation quality of details such as body outline color and avoid the loss of small target vehicles; the L1 regularization is introduced to control the sparsity of generator model channels to realize channel pruning, distillation loss is added in the pruning process to dynamically adjust the number of channels of the student network, constructing the lightweight model NightvisionGAN_light to improve the detection speed; the proposed image enhancement method is jointly trained with YOLOv3 to improve the detection accuracy. Verified by the BDD dataset, the proposed algorithm can effectively reduce the leakage rate and false detection rate of nighttime vehicle detection.