SAL-YOLO-DeepSeek: a lightweight real-time detection and LLM-driven decision framework for intelligent escalator safety monitoring

Abstract

To address the challenges of high passenger density, severe object occlusion, and limited computational resources in escalator scenarios, this paper proposes SAL-YOLO, a lightweight detection model, and develops a safety warning system. The model is optimized in multiple dimensions based on the YOLOv8n architecture: (1) A lightweight backbone network, StarNet, is designed by integrating depth-wise separable convolutions and residual connections to improve feature extraction efficiency. (2) A Transformer-based Adaptive Image Feature Integration (AIFI) module is integrated to capture global contextual dependencies through a self-attention mechanism. (3) The C2f_star feature fusion module is designed to enhance the discriminative power of local features in scenarios with occlusion and small objects. (4) A Lightweight Shared Convolution (LWSC) detection head is proposed that uses shared convolution and a dynamic stride mechanism to optimize computational load. Experiments on a custom-built dataset show that SAL-YOLO achieves 92.6% Precision (P), 95.4% mAP@50, and 79.9% mAP@(0.5:0.95). The computational complexity and parameter count are reduced by 47.1% and 50%, respectively, compared to the baseline model YOLOv8n. Furthermore, by introducing a knowledge distillation framework with a Similarity-Preserving (SP) feature distillation strategy, the model’s performance improved to 94.6% P and 80.6% mAP@(0.5:0.95), achieving an inference speed of 500 FPS. Building on the aforementioned model, an innovative safety warning system is developed by integrating the DeepSeek large language model, establishing an end-to-end mechanism of ’behavior recognition - risk assessment - response plan generation’. When abnormal behaviors, such as falls or bending over, are detected, the system dynamically generates graded response strategies and triggers multi-modal alerts. Empirical validation shows that the system is highly robust in complex scenarios, such as varying illumination and dense occlusions, offering both a theoretical methodology and an application paradigm for intelligent public safety surveillance.