LIGHTWEIGHT REAL-TIME OBJECT DETECTION FOR ASSISTIVE VISION SYSTEMS USING YOLOV3-TINY AND OPENCV

Abstract

This paper presents the development of a lightweight, real-time object detection system specifically designed to assist visually impaired individuals in identifying surrounding objects, thereby improving navigation and personal safety. Utilizing the YOLOv3-tiny model integrated with OpenCV, the system captures and processes real-time webcam video streams to recognize multiple object categories. Non-Maximum Suppression (NMS) and configurable confidence thresholds (0.4, 0.6, 0.8) are employed to optimize the balance between detection precision and recall. These threshold values were selected based on prior research and common practice in YOLO-based studies to provide a representative range for performance evaluation. Achieving an average of 25–30 frames per second (FPS) on standard hardware, the system demonstrates robust detection capabilities, even in moderately complex scenes. Comparative analysis with other lightweight models highlights YOLOv3-tiny’s advantage in speed and accuracy balance, making it suitable for mobile and embedded deployment. The results indicate potential for integrating audio feedback and adaptive thresholding in future versions to further enhance accessibility for visually impaired users.