Research On Vehicle Detection Method At Nighttime Based On Joint Generative Adversarial Network

Vehicle detection based on visual image is one of the key problems in the field of computer vision.Although vehicle detection has achieved good results in daytime,there are still many research difficulties at nighttime environment.Existing night vehicle detection methods mainly detect vehicles by detecting headlights or taillights.However,these features are adversely affected by the complex lighting environment.According to the existing difficulties of night vehicle detection,a method based on joint generative adversarial network(GAN)for nighttime vehicle detection is proposed in this thesis,which can effectively enhance the night vehicle visual information and improve the vehicle recall rate and reduce the false detection.The main work of this thesis is as follows:(1)A feature translate-enhancement generative adversarial network for vehicle detection(Fte Gan Vd)cascade network is proposed to solve the complex lighting in night environment.Firstly,Fte Gan Vd is composed of two modules: 1)Feature enhancement module based on Cycle GAN: The nighttime image is transform into daytime image by the generator of enhancement module to improve the brightness and enhance the contrast between the vehicle and background.Then,the multi-scale features of night and day image are fused to enhance the image feature;2)Vehicle detection module: The existing object detection network is cascaded with the enhancement module as a detector.The enhanced feature is used as the input of the detection module for vehicle detection.Secondly,in order to improve the attention of the network to vehicle,we propose an improvement to the loss function of the generator.A weight mask is generated by the position of vehicle to adjust the loss proportion of vehicle and background,which make the network more focused on the conversion of the vehicle.(2)According to the real-time requirements of practical applications,this thesis proposes a multi-task network for night image transformation and vehicle detection.Firstly,the image transformation and vehicle detection share the same backbone network used to extract the similarity information of night images.Then night to day image transformation and nighttime vehicle detection are performed at the same time by two sub-networks.Finally,the multiscale features of the transformation task and the detection task are fused to enhance the robustness of the vehicle features to improve the accuracy of the detection task.To verify the effectiveness of the proposed method,experiments are conducted on the public dataset Berkeley Deep Drive(BDD)and the private dataset in this thesis.Comparing with the existing object detection network,the average precision of our method is improved on different datasets.Comparing the recall and precision,the experimental results show that the method in this thesis can effectively improve vehicle recognition and reduce the false detection rate while improving the vehicle recall rate.