Occlusion-aware Perception and Planning for Automated Vehicles

Abstract

The perception system is a key component of automated vehicles, as it relies on onboard sensors to gather information. However, the system faces challenges due to occlusions that obstruct visibility. Safely navigating in highly occluded scenarios remains a significant obstacle for automated vehicles. In this dissertation, we present a systematic approach to address this issue by preparing automated vehicles for fully occluded areas on the road.

To effectively model the occluded areas, we introduce a joint object detection and semantic segmentation algorithm. This helps in acquiring critical environmental information with increased efficiency, enabling automated vehicles to make decisions in the presence of occlusions. In tandem, we propose a semantic 3D mapping framework that efficiently identifies occlusions, which feeds into a layered 2D map, essential for planning, and contains the occlusion data. The experiments in SemanticKITTI dataset demonstrated that the proposed perception algorithms can generate a semantic grid map of the environment and identify the occluded grids efficiently and effectively.

To tackle the occlusion problem and produce a safe plan for the automated vehicle, an occlusion-aware object management system is introduced to generate virtual road users for the planning algorithm, and near-optimal trajectories are solved using a sampling-based method while taking the presence of virtual objects into account. The experiments in a 2D toy driving environment showed the proposed planner can achieve better safety against baseline approaches while maintain a reasonable passing speed in several challenging testing scenarios.

Furthermore, the effectiveness of the proposed perception and planning framework is validated in both the Carla simulator and the physical Mcity testing facility, demonstrating the effectiveness of the proposed architecture and its superior safety performance over baseline approaches. Besides, a modular AV stack is described to guide the integration of the proposed perception and planning framework in the experiments.

Validation experiment results show that the proposed framework results in a reduced crash rate in comparison to several baselines, including the renowned open-source AV framework, Autoware. Notably, these outcomes were realized without needing a high-definition map with road geometry definitions.