Empowering the Visually Impaired: YOLOv8-based Object Detection in Android Applications

Abstract

This research helps visually impaired individuals enhance their quality of life using object detection technology. Vision is a fundamental sense that lets people interact and navigate through their surroundings. Earlier, object detection was implemented with methods like Sliding Window, and Haar Cascades which lacked the accuracy and efficiency that are necessary for a real-time environment. This study proposed an object detection system specifically defined for visually impaired individuals using the YOLOv8 model because YOLO is known for its real-time performance, accurate detection, and classification capabilities, ensuring instant and accurate assistance in real-time environments. Additionally, YOLO’s robustness in detecting objects in various scenarios makes it a reliable solution for aiding visually impaired individuals in navigating their surroundings effectively. The novelty of the study lies in converting the YOLOv8 model into TensorFlow Lite, making it easier to deploy on Android platforms and widening its accessibility to a larger user population. The application of this object detection system extends to various aspects of everyday life, promoting inclusivity and accessibility for visually impaired individuals and bridging the gap between cutting-edge research and practical assistive technologies with the help of support on edge devices and auditory feedback. The performance metrics are evaluated on multiple models to understand the performance of the proposed model on different training parameters such as model size and input image resolution. The proposed model with an input image size of 640X640 resolution and Nano-sized model achieved mAP50 and mAP50-95 of 79.9% and 57.1% respectively with inference time of 350ms on average. For the same resolution the proposed model is trained on Medium-sized model and attained mAP50 and mAP50-95 as 80.5% and 60.9% respectively with inference time of 1800ms average. For an image size of 256X256 resolution and Nano-sized model the proposed model achieved mAP50 and mAP50-95 of 74.4% and 52.9% respectively with inference time of 55ms on average. From results it is observed that the model accuracy increases as the model size and input resolution is high but it increases the inference time. Whereas the model with less inference time lower input image resolution gives most efficient results but also reduces accuracy.

Topics & Concepts