Performance Assessment of Fine-Tuned Barrier Recognition Models in Varying Conditions

Walking has been widely promoted by various medical institutions as a major contributor to physical activity that keeps people healthy. However, pedestrian safety remains a critical concern due to barriers present on sidewalks, such as bins, poles, and trees. Although pedestrians are generally cautious, these barriers can pose a significant risk to vulnerable groups, such as the visually impaired and elderly. To address this issue, accurate and robust computer vision models can be used to detect barriers on pedestrian pathways in real-time. In this study, we assess the performance of fine-tuned egocentric barrier recognition models under various conditions, such as lighting variations, angles of view, video frame rates and levels of obstruction. In this context, we collected a dataset of different barriers, and fine-tuned two representative image recognition models, assessing their performances on a set of videos taken from a predefined route. Our findings provide guidelines for retaining model performance for applications using barrier recognition models in varying environmental conditions.