Based on YOLO26L and Attention-LSTM Intelligent Public Space Pedestrian Flow Sensing Platform

Abstract

Public space pedestrian flow management faces three core statistical challenges: accurate crowd counting, prediction distortion in sparse time series, and quantitative assessment of congestion levels. Using a university teaching building as the experimental scenario, this paper constructs an intelligent pedestrian flow analysis framework that integrates computer vision and statistical modeling. At the detection layer, a fine-tuned YOLO26L model combined with a 3-second cooldown mechanism reduces the line-crossing counting error rate to 5.71%. At the time-series prediction layer, to address the failure of MAPE (Mean Absolute Percentage Error) when zero-value frames exceed 50%, Symmetric Mean Absolute Percentage Error (SMAPE) is introduced as the core evaluation metric, and an Attention-LSTM model incorporating 35-dimensional temporal cycle, calendar semantic, lag, and rolling statistical features is constructed, achieving a test-set SMAPE of 35.06%—approximately 44% better than the baseline LSTM (Long Short-Term Memory). At the congestion assessment layer, a triple quantitative system combining a pixel density model, the M/M/c Erlang-C queuing model, and the SFPE (Society of Fire Protection Engineers) evacuation engineering model is established, realizing a complete statistical inference pipeline from density sensing to tiered early warning. The system achieves a detection mAP50 of 0.993 with an inference speed of 110.1 ms/frame, with all statistical metrics meeting preset targets, demonstrating strong potential for broad deployment.