LiCaNet: Further Enhancement of Joint Perception and Motion Prediction Based on Multi-Modal Fusion

Abstract

The safety and reliability of autonomous driving pivots on the accuracy of perception and motion prediction pipelines, which reckons primarily on the sensors deployed onboard. Slight confusion in perception and motion prediction can result in catastrophic consequences due to misinterpretation in later pipelines. Therefore, researchers have recently devoted considerable effort towards enhancing perception and motion prediction models. However, targeting pixel-wise joint perception and motion prediction using different sensor modalities are often ignored. In this paper, we push performance even further by leveraging a multi-modal fusion network. We propose a novel LIDAR Camera Network (LiCaNet) that achieves accurate pixel-wise joint perception and motion prediction in real-time. LiCaNet expands on our earlier fusion network by incorporating a camera image into the fusion of LIDAR sourced sequential bird’s-eye view (BEV) and range view (RV) images. We present a comprehensive evaluation using nuScenes dataset to validate the outstanding performance of LiCaNet compared to the state-of-the-art. Experiments reveal that utilizing a camera sensor results in a substantial gain in perception and motion prediction. Moreover, most of the improvements achieved fall within the camera range, with the highest registered for small and distant objects, confirming the significance of incorporating a camera sensor into a fusion network.