Ensuring the safety of pedestrians and cyclists referred to as vulnerable road users (VRUs) is a central challenge in intelligent transportation systems (ITS), especially in urban environments where visual occlusion, behavioral unpredictability, and data scarcity pose significant barriers to real-time detection. This paper introduces STEP-SAFE, a novel few-shot learning framework optimized for TinyML platforms that enables accurate and efficient VRU recognition and alert generation with minimal supervision and extremely low computational overhead. The proposed model utilizes a prototype based meta learning architecture in combination with a lightweight convolutional encoder to support rapid generalization from as few as one to five labeled samples per class. To simulate embedded deployment, we integrate quantization aware training and sparsity based pruning, reducing the model size to under 300 KB and enabling inference latency below 64 milliseconds on micro-controller class hardware. Additionally, STEP-SAFE incorporates a dynamic risk aware alert mechanism that converts classification confidence into real-time safety decisions. The model is evaluated across multiple few-shot configurations (1-shot to 10-shot, 3-way to 5-way) and benchmarked against state-of-the-art methods. Moreover, results show that STEP-SAFE achieves up to 85.2% accuracy and 0.85mAP while requiring significantly fewer computational resources and support samples. Confusion matrix and episodic performance analyses confirm its robustness in distinguishing visually similar and occlusion prone VRU behaviors. By combining compact model design, few-shot adaptability, and interpretable alert logic, STEP-SAFE represents a deployable and intelligent solution for future ready ITS safety infrastructure.
Khalid et al. (Wed,) studied this question.