基于深度学习的障碍物检测与深度估计

提出了一种针对交通场景的基于深度学习的障碍物检测与深度估计方法。该方法对现有的YOLOv3模型进行改进，使用DenseNet网络代替原网络尺度较小的传输层，得到一种新的障碍物检测模型Dense-YOLO。然后采用立体匹配模型PSMNet得到双目图像的视差图，根据双目测距原理对被测目标深度进行估计。在KITTI数据集和实际交通场景中的实验结果表明，与YOLOv3模型相比，Dense-YOLO模型有效地提高了交通场景中障碍物检测的可靠性和正确率，对轿车、行人、骑行者和卡车这4类障碍物检测的平均精确率（average precision, AP）提高了3%～5%，平均精确率均值（mean average precision, mAP）提高了约4%。障碍物深度估计结果与真实值的平均相对误差约为3%。

Obstacle detection and depth estimation using deep learning approaches

An obstacle detection and depth estimation method based on deep learning for traffic scenarios was proposed. The method modifies the existing YOLOv3 model by replacing its small-scale transmission layer with DenseNet network, and obtains a new obstacle detection network Dense-YOLO. Then, the disparity map of binocular images was obtained by using the stereo matching network model PSMNet, and the depth of detected obstacles was estimated according to the binocular ranging principle. Extensive experiments were conducted on the KITTI dataset and the actual traffic images, and the results show that Dense-YOLO effectively improves the reliability and accuracy of the obstacle detection in traffic scenarios compared to YOLOv3. The average precision (AP) on the four-class of obstacles including car, pedestrian, cyclist and truck is increased by 3% to 5%, and the mean average precision (mAP) is increased by about 4%.The average relative error between the estimated depth of the obstacle and the true value is about 3%.