Abstract: With the continuous development of target detection technology,the vehicle detection system for road scenes has been widely used in the field of automatic driving. Compared with the traditional target detector, though the target of vehicle detection is relatively simple, two major problems need to be solved. First, the characteristics which provided to the detector are usually incomplete in some complex road scenes, and other problems such as occlusion and deformation will occur. Second, it is necessary to estimate the distance of different vehicles to ensure the car can make timely evasive action in the process of automatic driving, which means it needs depth estimation of the target area of the image. Aiming at these two problems, a vehicle detection system based on anti⁃sample generation and depth map reconstruction was proposed. A confrontation network was designed for the pre⁃training target detection network called Faster⁃RCNN, which was used to generate a large number of samples during the training process, and train the vehicle detector with these samples. According to the detection results, the vehicle distance is estimated to inform the system to make evasive action in time through the reconstruction of 3D scene and camera pose recovery depth map. The experimental results show that this detection system can improve the detection effect and estimate the distance of the target vehicle without increasing the data training sample.

Key words: Object detection, Adversarial learning, Depth estimation, 3D scene sparse recovery

摘要： 随着目标检测技术的不断发展，用于道路场景的车辆检测系统在自动驾驶领域得到了广泛应用。与传统的目标检测器相比，车辆检测的目标比较单一，但同时需要解决两大问题，一是在复杂的道路场景中，提供给检测器的车辆特征通常是不完整的，会出现遮挡和形变等问题；二是在自动驾驶过程中，需要对不同车辆的距离做出估计才能保证智能车及时地做出规避动作，即对图像的目标区域进行深度估计。针对这两个问题，提出了基于对抗样本生成与深度图重建的车辆检测方法。为预训练目标检测网络Faster⁃RCNN设计一个对抗网络，用于在训练过程中产生大量的训练样本，并利用这些样本对车辆检测器进行训练；根据检测结果，通过重建3D场景与相机位姿恢复深度图，对车辆的距离做出估计，以通知系统及时做出规避动作。实验结果表明，在不增加数据训练样本的情况下，该检测系统可以较好地提升车辆检测效果及估计目标车辆的距离。

关键词: 目标检测, 对抗学习, 深度估计, 　3D 场景稀疏重建