Collision-Free Path Generation for Teleoperation of Unmanned Vehicles

Teleoperation, or remote driving, constitutes a crucial transitional phase toward the widespread adoption of fully autonomous vehicles. Nevertheless, to enable seamless real-time teleoperation, it is imperative to address the time delay between the driver and the vehicle. Collision-free path generation has emerged as a vital technique facilitating both teleoperation and autonomous driving, particularly in high-level path planning for vehicles. In the context of real-time teleoperation, a generated collision-free path serves as a valuable guide for the teleoperator, effectively mitigating the impact of time delay. In this research, we present a framework dubbed dual transformer network (DTNet), designed to cater to the needs of teleoperation by addressing road scene understanding. The proposed DTNet employs two transformer-based networks to effectively segment the road free space and detect road objects. Additionally, we introduce an innovative fusion mechanism that leverages the combined information from both networks to predict a collision-free path. The efficacy of the DTNet is extensively evaluated using a large-scale BDD100k dataset, substantiating its superior performance in road free space segmentation and road object detection tasks. Remarkably, DTNet achieves a mean intersection over union score of 83.89% for road free space segmentation and an impressive mean average precision score of 34.20% for road object detection. The experimental findings affirm the effectiveness of the DTNet framework in addressing the challenges of road scene understanding, making it a promising solution to provide a robust and efficient approach for collision-free path generation, with broader implications for the advancement of autonomous driving technologies.