Adaptive Feature Fusion Based Cooperative 3D Object Detection for Autonomous Driving

In this paper, we focus on the collaborative 3D object detection problem in autonomous vehicle systems in which autonomous vehicles can improve their detection accuracy by aggregating the information received from spatially diverse sensors through wireless links. We propose a novel adaptive feature fusion based cooperative 3D object detection framework, which consists of feature transformation networks and an improved region proposal network. The framework learns to fuse features from different views to improve object detection accuracy on the autonomous vehicle. To evaluate the proposed method, we build a new synthetic dataset created in two driving scenarios (a Roundabout and a T-junction). Experiment analysis and results demonstrate that the proposed adaptive feature fusion approach performs better than two baseline approaches in terms of detection accuracy.