Towards Compact Autonomous Driving Perception with Balanced Learning and Multi-sensor Fusion
Researchers have developed a new compact deep learning model for autonomous driving that can perform multiple perception tasks simultaneously. This model integrates semantic segmentation, depth estimation, LiDAR segmentation, and bird's-eye view projection in a single forward pass. It utilizes an adaptive loss weighting algorithm to address imbalanced learning across tasks and fuses data from RGB cameras, dynamic vision sensors, and LiDAR for a comprehensive environmental understanding. The model demonstrates superior performance with fewer parameters, leading to faster inference and reduced GPU memory usage, with consistent results across simulation and real-world datasets. AI
IMPACT This compact model could enable more efficient and capable perception systems for autonomous vehicles, potentially reducing hardware costs and improving real-time performance.