Object Recognition Method Commonly Usable for LIDARs With Different Vertical Resolution

In order to recognize objects with high robustness using low resolution LIDAR, we proposed a two-stage deep neural network method that integrated independent results for each scan line and calculated the final recognition result. In this paper, we optimize the former stage with unsupervised learning and add a pooling layer between the former stage and the latter stage for the purpose that the proposed method is commonly usable for the LIDARs with the different vertical resolution. We classified the LIDAR data of pedestrians, bicycles, motorbikes, cars and other objects, and evaluated the accuracy of the improved proposal method.