Multi-Scene Dataset and Object Detector for Outside Blind Individual Identification

Haotian Ji, Israel Mendonça, Masayoshi Aritsugi

Abstract: In the field of Computer Vision (CV), assistive navigation systems for visually impaired individuals have garnered significant attention in recent years. However, most existing solutions rely on wearable devices or mobile platforms, which often face limitations in cost, deployment, and robustness in complex outdoor environments. This paper proposes a practical approach to public space recognition for the visually impaired. The proposed recognition method can be widely applied to existing public video surveillance systems. We created a specialized image dataset tailored for recognition by visually impaired individuals. At the same time, for recognition tasks involving nighttime environments and partially occluded targets, we propose a composite framework that integrates various grouped convolutions and image enhancement networks. Compared to the original baseline detection models, our models outperform on our created dataset by 1.5% average precision (e.g., from 94.1% to 95.6% for YOLOv8x), while reducing parameters by up to 35.7% (e.g., from 56.9M to 36.6M for YOLOv11x). Furthermore, our models also achieve over 0.8% AP and a parameter reduction exceeding 10% compared to the original baseline models on ExDARK dataset.