Paper Details
Abstract
Autonomous navigation in dynamic environments presents significant challenges due to the inherent limitations of individual sensing modalities. While 2D LiDAR excels in geometric mapping but fails to capture semantic information and height variations, RGB-D cameras provide rich semantic context but suffer from lighting sensitivity and range limitations. This paper introduces a novel predictive occupancy mapping framework that leverages multi-modal sensor fusion with confidence-aware motion prediction for enhanced robotic navigation safety. Our approach integrates YOLO-based object detection with DeepSORT tracking to establish persistent object identities, enabling sophisticated motion pattern analysis through three distinct kinematic models: Stopped, Linear, and Curve motion. The system employs a three-tiered confidence fusion strategy (Camera_Only, LiDAR_Enhanced, LiDAR_Validated) that adaptively weights sensor contributions based on spatial-temporal consistency. Another improvement lies in the occupancy prediction grid generation, which combines both current object locations and predicted future locations encoded at different occupancy levels. The framework demonstrates computational efficiency through incremental grid updates that exploit persistent track identities, achieving real-time performance on resource-constrained platforms.