MLRDv2: A Dataset for Improving Micromobility Safety via Attention-Integrated Compact CNN Models

Micromobility offers a sustainable alternative to traditional transportation but lacks clear safety regulations. Existing sensor-based solutions for micromobility safety are often imprecise, expensive, or resource-intensive, making them unsuitable for constrained environments. AI-based lane detection techniques hold potential, but most rely on image segmentation, which is computationally demanding. A significant challenge is the absence of dedicated image datasets for micromobility, as current datasets primarily focus on autonomous driving and do not capture the unique perspectives of micromobility vehicles. To address this, we introduce the Micromobility Lane Recognition Dataset (MLRD), which enables real-time lane identification to regulate rider behavior. Using MLRD, we evaluate the effect of channel and spatial attention mechanisms on compact convolutional neural networks (CNNs). Our findings show that combining channel and spatial attention improves CNN performance by enabling better focus on important features. MobileNet V2, with integrated attention mechanisms, achieved the highest precision and F1 scores, while MobileNet V3 maintained strong performance with fewer parameters. To meet the growing demand for micromobility, we also present MLRDv2, an improved dataset featuring more diverse scenarios. Testing on MobileNet V2 and V3 Large models showed a 4% performance boost compared to results from MLRD V1, demonstrating the dataset’s effectiveness.