Semantic-based Transitional Grid Maps for Autonomous Driving

Sammanfattning

This work investigates the integration of semantic perception into probabilistic dynamic grid-based mapping for autonomous driving applications. The focus is on improving the classification of obstacles as static or dynamic within the Transitional Grid Map (TGM) framework, where conventional approaches rely primarily on LiDAR-based geometric and motion cues. In real-world urban environments, such cues are often insufficient due to temporary object immobility, occlusions, and sensor noise, leading to degraded classification performance. Accurate static–dynamic classification is critical, as errors directly affect downstream tasks such as localization, prediction, and motion planning. Despite advances in probabilistic mapping and vision-based semantic perception, these approaches are typically developed independently, and their combined effect within unified frameworks remains insufficiently explored. To address this limitation, a semantic-enhanced extension of the TGM framework is proposed. The approach integrates camera-based object detection using a You Only Look Once (YOLO) deep learning model into the probabilistic mapping pipeline through a multi-modal sensor fusion strategy. Semantic labels are projected onto LiDAR point clouds, clustered, and incorporated into a semantic inverse sensor model, which influences the probabilistic allocation of grid cells to static and dynamic states while preserving the underlying Bayesian formulation. The method is evaluated using real-world urban driving data collected in Stockholm, under controlled comparison with a LiDAR-only baseline. The results demonstrate substantial improvements, with overall accuracy increasing from 0.64 to 0.93 and recall from 0.59 to 0.92. The gains are particularly significant for context-dependent objects, such as temporarily stationary vehicles. These results demonstrate that integrating semantic perception into probabilistic grid-based mapping significantly improves robustness in dynamic environments, enabling more reliable static–dynamic classification and providing a stronger foundation for downstream autonomous driving tasks such as prediction and motion planning.