LDD-Track: An energy-efficient deep reinforcement learning framework for multi-subject tracking in mobile crowdsensing

Abstract

Multi-subject tracking in Mobile Crowdsensing Systems (MCS) is a challenging task due to dynamic mobility, limited energy resources, and the need for real-time decisions. Traditional models like Kalman Filters and Hidden Markov Models struggle in such conditions, while Transformer-based deep learning methods offer high accuracy but are too computationally demanding for mobile use. Unlike previous studies that focus on one-to-one or collaborative group tracking, which often lack scalability and adaptability to real-world complexities, we propose LDD-Track, a novel multi-subject tracking framework that integrates Long Short-Term Memory (LSTM) networks with an adaptive attention mechanism, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and Deep Q-Network (DQN)-based user allocation. The LSTM model, enhanced with attention mechanisms, dynamically assigns weights ${\alpha }_t$ to past trajectory points, filtering noise and improving prediction accuracy. The DBSCAN clustering technique effectively groups subjects based on predicted movement, optimizing resource allocation and reducing computational overhead. The DQN-based user assignment strategy models resource optimization as a Markov Decision Process (MDP), leveraging the Q-value function $Q(s_t,a_t)$ to ensure adaptive and energy-efficient user allocation. Extensive experiments on the Taxi Mobility in Rome dataset demonstrate the superiority of LDD-Track. The framework achieves a 51 % reduction in energy consumption, a 39 % increase in Coverage Completion Rate (CCR), and a 9.7 % improvement in resource allocation efficiency compared to state-of-the-art methods. These findings validate the effectiveness of integrating attention-based prediction and deep reinforcement learning in large-scale, real-time MCS environments.