Attention Module in YOLO-Based Object Detection Method for Autonomous Smart Wheelchair Room Navigation System

People with visual impairment have difficulty in navigating a room based on text in room nameplate. Recognizing room nameplate in real environments with computer vision approach is challenging, because the system has to detect the room nameplate objects. Detection result can be improved by adding attention module in deep learning architecture, however model complexity also need to be observed to prevent accident caused by slow response from the system. Based on that problem, we proposed to compare the effect of Coordinate Attention (CA), Convolutional Block Attention Module (CBAM), and Shuffle Attention (SA) on YOLOv8 model. Based on our findings, CA module shows positive result on accuracy of 99,50%, precision of 1, and f1-score of 0,997. As for CBAM has reduced accuracy to 96,15% from original YOLOv8n that has 98,52% accuracy, this may occur because the placement of the CBAM module before detection head is not suitable in the case of room name plate detection. Meanwhile, SA module has the least number of parameters and model size increase of additional 73.896 parameters and 154 KB, respectively. Our findings enrich insight in improving object detection method for autonomous smart wheelchair room navigation system.