DEVELOPMENT OF AN OPTIMIZED YOLO-BASED REAL-TIME VISION SYSTEM FOR OBJECT DETECTION IN INDUSTRIAL CLEANING ROBOTS

Industrial cleaning robots operating in dynamic and cluttered environments require robust, accurate, and low-latency object detection systems to ensure safe navigation, efficient task execution, and collision avoidance. Traditional vision-based detection methods often fail to meet real-time performance constraints when deployed on embedded robotic platforms with limited computational resources. This research presents the development of an optimized YOLO-based real-time vision system for object detection in industrial cleaning robots. The proposed system leverages advanced variants of the YOLO (You Only Look Once) deep learning architecture, selected for its single-stage detection capability and high inference speed. Model optimization techniques including network pruning, quantization, input resolution tuning, anchor box recalibration, and hardware-aware training were implemented to balance detection accuracy and computational efficiency. Transfer learning was employed using pre-trained weights, followed by fine-tuning on a domain-specific dataset consisting of industrial floor environments with obstacles such as debris, cables, tools, liquid spills, and human presence. To ensure real-time deployment, the system was integrated with embedded GPU hardware and optimized using parallel processing and model acceleration frameworks. Performance evaluation was conducted using standard metrics including mean Average Precision (mAP), precision-recall curves, F1-score, and inference latency (frames per second). Experimental results demonstrate that the optimized YOLO model achieves high detection accuracy while maintaining real-time performance under industrial lighting variations, occlusions, and dynamic object movement. The developed vision system significantly enhances robotic perception capabilities, enabling improved obstacle avoidance, adaptive cleaning path planning, and operational safety. The study contributes a scalable and deployable framework for integrating lightweight deep learning-based object detection into autonomous industrial service robots. Future work may explore multimodal sensor fusion, edge-AI deployment optimization, and continual learning for long-term industrial adaptability.