Personnel Detection via Reinforcement Learning-Based Dynamic Parameter Optimization with Vehicle-Mounted Ultra-Wideband

With the development of intelligent cockpit technology, detecting the presence of occupants inside the vehicle has become a core requirement for enhancing driving safety and interaction experience. However, existing solutions based on vision, millimeter-wave radar, or dedicated sensors face limitations due to environmental interference, high cost, privacy concerns, or hardware complexity. The widespread adoption of Ultra-Wideband (UWB) digital keys has opened up new opportunities for personnel detection. However, current UWB solutions suffer from issues such as reduced detection sensitivity caused by statically configured communication parameters and the inability to accurately distinguish between humans and objects. This paper proposes a UWB-based personnel detection system UAware, which achieves efficient sensing through reinforcement learning-driven dynamic parameter optimization and a two-stage detection architecture. At the parameter optimization level, a Markov Decision Process (MDP) model is constructed, and an improved Hierarchical Proximal Policy Optimization (PPO) algorithm is proposed to address high-dimensional nonlinear coupled parameter configuration problems. At the detection level, a novel two-stage inference process is introduced, which enables millisecond-level detection through link packet loss rate monitoring and distinguishes between personnel and objects by leveraging the fluctuation characteristics of the preprocessed Channel Impulse Response (CIR). Experiments demonstrate that UAware achieves a detection accuracy of 93% in dynamic vehicular scenarios, a 17.5% improvement over traditional methods, with latency reduced to the millisecond level.