FusionGAN-Detection: vehicle detection based on 3D-LIDAR and color camera data

At present, most of the deep learning target detection methods based on multimodal information fusion are integrated, which makes the fusion image quality cannot be directly controlled. It is not conducive to strengthening the target detection of the network in principle. A multimodal information fusion detection method based on generative countermeasure network (FusionGAN-Detection) is proposed, which is composed of GAN and a target detection network. Aiming at the uncontrollability and blindness of existing information fusion detection algorithms, the new method introduces generative countermeasure network for information fusion. It uses loss function and dual discriminator to provide controllable guidance for generator. In the process of information fusion using GAN, the loss function of traditional thought can extract the information which is beneficial to target recognition to the maximum extent, and avoid the loss of channel information. The detector acts as a discriminator during the training process to guide image fusion and promote the improvement of image quality. Meanwhile, it acts as a target detector for target detection during testing. In order to verify the effectiveness of the method, KITTI data sets are used for training and testing. The experimental results show that new method is better than the existing advanced methods in AP.