3D-AOCL: Analytic online continual learning for imbalanced 3D point cloud classification

Abstract

Recent autonomous driving systems heavily rely on 3D point cloud data collected from multiple sensors for environmental awareness and decision-making. However, it is unrealistic to expect the autonomous driving system to recognize all road environments and handle every traffic situation. Models for autonomous driving need to be updated in real time in order for the system to adapt to more situations. This is where online continual learning becomes crucial. Online continual learning is an important method in the field of autonomous driving, as it enables models to update their parameters with streaming input data for adapting to new environments and conditions. Online continual learning in the field of autonomous driving faces several challenges: inefficient data fusion, catastrophic forgetting, insufficient computational resources, violation of road privacy and categories imbalance. To tackle these challenges, we propose an Analytic Online Continual Learning method for 3D Point Cloud Classification (3D-AOCL). This approach utilizes Analytic Learning to update parameters and integrates a feature fusion module along with a category balancer to address the above issues. It is capable of fusing data in feature level, balancing samples across various categories and updating parameters by calculating the analytical solution. We have validated our method on the vehicle side, the infrastructure side, and vehicle-infrastructure cooperative data on the V2X-Seq dataset. The experimental results demonstrate that our model effectively addresses key issues in online continual learning for autonomous driving systems, outperforming other models by approximately 4.00% to 6.00% in AMCA scores while only keeping 0.75% trainable parameters.