Wheel Center Detection From Point Cloud Based on Recurrent Neural Networks

In the intelligent parking lot, the robot has to get the accurate position of the wheels in the process of automatic parking. The traditional image processing algorithm is not proper to be employed in this application because of sensitivity to the noises. Originating from natural language processing, RNN has now been used in the computer vision for its robustness and excellent performance. This paper proposes a method to position vehicle wheels from point cloud based on the RNN model. The experimental results show that the prediction accuracy of the bidirectional RNN is obviously higher than that of the unidirectional RNN. In a small-scale training set, the effect of BiLSTM/BiGRU is close, while the training speed of BiGRU is faster, which can be used in the wheel detection system.