Research on Visitor Flow Recognition and Regulation Optimization in Scenic Areas Based on Deep Transfer Learning

To address sparse samples, poor scenario adaptability, and imprecise control strategies in scenic-area visitor-flow behavior recognition, this paper studies the application of deep transfer learning to visitor-flow recognition and regulation optimization. Integrating visitor behavior science, deep transfer learning, and crowd dynamics theory, the study establishes a technical framework of “data collection-feature transfer-behavior recognition-control optimization, ” and defines core visitor behavior types and regulation objectives. To solve data scarcity in small-sample scenarios, multi-source data including video surveillance, GPS trajectories, and ticketing data are fused to construct the STBD- 2024 scenic-area visitor-flow dataset, covering six typical behavior categories. A deep transfer learning model based on improved ResNet50 and domain adaptation is then designed, improving cross-scenario behavior recognition accuracy through pre-training and fine-tuning. Finally, behavior-recognition results are combined with scenic-area spatial layout features to propose a three-dimensional regulation strategy of zonal regulation, dynamic diversion, and precise early warning. Simulation experiments and field validation quantify the effectiveness of the model and regulation strategy.