DRIFT: A Dynamic Crowd Inflow Control System Using LSTM-Based Deep Reinforcement Learning

Crowd management plays a crucial role in improving travel efficiency and reducing potential risks caused by overcrowding in large public places. Crowd control at entrances is a common way in our daily life to avoid overcrowding, but nowadays the control of crowd inflow at the entrances of public places mainly relies on manual operation. In this article, we intend to propose a dynamic crowd inflow control system (DRIFT) to avoid risks of overcrowding and improve the throughput of public places. First, we formulate an optimization problem that maximizes throughput by adjusting the crowd inflow rate of each entrance in the public place. Through mathematical analysis and related proofs, we introduce a baseline for the aforementioned problem that can calculate the upper bound of static inflow rate. With this baseline, we can easily measure the performance of other dynamic inflow control algorithms. Second, we treat the proposed optimization problem as a real-time decision-making problem, and further propose the DRIFT system based on deep reinforcement learning to address it. Specifically, the strategy of DRIFT is a basic actor-critic framework adapting a shared long short term memory (LSTM) layer to extract scene feature information. Third, we train it through proximal policy optimization (PPO) to improve learning performance. The environment for experiments is a crowd simulation model of OpenAI Gym structure based on real scene data from the 1F floor of the Chengdudong Railway Station and Xizhimen Railway Station. In comparison experiments and ablation experiments, the strategy of our DRIFT outperforms all other comparison strategies, including the most recent strategy using reinforcement learning, in term of system crowd throughput and robustness.