未知低障害物回避のための三次元幾何情報とDeep Neural Networkによる立体物検出

抄録

In this study, we developed a new method to detect unknown low-height obstacles using 3D point clouds from stereo cameras. Conventional semantic segmentation methods using a depth image by Deep Neural Network (DNN) can detect road surfaces with a high accuracy. However, it is difficult to detect unknown low-height obstacles not included in the training data. Methods that use geometric information such as normal and height face difficulty to find objects with a surface that parallels to the road surface and low objects, respectively. Therefore, we deal with the difficult problem of detecting unknown low-height obstacles. To solve the problem, we focused on the difference in difficult detection between DNN and geometric methods. Based on the confidence from the output of DNN, we help difficult obstacle detection for DNN by using geometric information, and vice versa. When tested on a robot equipped with a stereo camera, the IoU, which indicates the detection accuracy of unknown obstacles, was improved by 18.1 percentage points compared to DNN. Moreover, our method enabled the robot to safely avoid three types of unknown low-height obstacles.