Improved algorithm for visually impaired individuals avoid obstacles based on YOLOv8

The use of object detection technology to assist visually impaired individuals in obstacle avoidance has gained increasing attention. However, in real-world scenarios, the scale variation of obstacles and the prevalence of small-sized obstacles pose significant challenges. This paper proposes a novel deep learning-based method for assisting the visually impaired in obstacle avoidance, which combines the YOLOv8 model, the EMA attention mechanism, WTConv technology, and an enlarged 160x160 object detection head to improve obstacle perception capabilities in complex environments. YOLOv8, as the baseline model for object detection, enables real-time perception, while the EMA attention mechanism selectively enhances important features, reducing noise and improving detection robustness under varying lighting conditions and occlusions. Furthermore, WTConv, through wavelet transformation, captures multi-scale spatial information, thereby enhancing the representation of fine details in low-resolution inputs and improving the accuracy of obstacle detection. Experimental results demonstrate that the proposed EMA-YOLOv8 model achieves an accuracy of 89.3% in complex real-world scenarios, outperforming the baseline model by 3.7%, highlighting its promising application in improving the safety and assistive technology for visually impaired individuals.