Deep Reinforcement Learning-Based Active Sensing on a Bicycle for Vehicle Tracking

초록

This article develops an intelligent active sensing algorithm using a low-cost on-bicycle sensor system to track approaching cars behind a bicycle. With the increasing popularity of electric-assist bicycles (e-bikes) and conventional bicycles, rider safety has become a growing concern. In this article, an inexpensive rotational laser sensor system is considered as an on-bicycle sensor system for measuring the distance to nearby cars. A new intelligent active sensing algorithm is developed based on deep reinforcement learning (DRL) combined with vehicle motion estimation algorithms. This algorithm enables the rotational laser sensor system to continuously focus on the front right corner position of a vehicle and to accurately estimate its motion. Extensive simulation studies are performed to evaluate the tracking performance of the developed algorithms. The simulations encompass various scenarios involving cars executing multiple maneuvers, such as straight driving, right and left turns, hard braking, and complete stops. Notably, the proposed algorithm exhibits superior performance in estimating the trajectory of vehicles performing dynamic maneuvers. In hard-braking scenario, the proposed method achieves a root-mean-square error (RMSE) of 0.1875 m for position and 1.0421 m/s for velocity, significantly outperforming the previously developed model predictive control (MPC)-based tracking method, which yields 0.2832 and 1.3072 m/s, respectively. Furthermore, after a complete stop, the proposed algorithm reduces the trajectory estimation error by approximately 53% and the velocity estimation error by 90% compared to the MPC-based method. These results demonstrate that the developed algorithm enables a new approach for cost-effective and robust vehicle tracking.

키워드