An Object Detection Method Based on Heterogeneous Lidar Point Clouds Using Federated Learning

3D Light Detection and Ranging (Lidar) is an essential sensor in intelligent unmanned systems. The surrounding traffic participants of an Unmanned Ground Vehicle (Ugv) are perceived from the Lidar point clouds with 3D spatial information. However, machine learning-based applications based on Lidar point clouds are strongly affected by the training data volume and the data distribution. As a result, a trained model from a limited dataset cannot meet the requirements of multiple Ugv equipped with heterogeneous Lidar in different scenarios. To solve these problems, this paper proposes an object detection method based on heterogeneous Lidar point clouds using federated learning (Fl). Local features are extracted from multiple Ugv agents while the server keeps updating a global feature. In order to improve the availability of the trained model, the loss of Maximum Mean Discrepancy (MMD) is used to compute the differences among global and local features. Experimental results indicate that our method not only improves the detection accuracy but also reduces communication rounds to obtain the final convergent model during the training process in heterogeneous scenarios.