Anindya Zulva Larasati, Fitri Utaminingrum, Rekyan Regasari Mardi Putri, M. Ali Fauzi
Many wheelchair users with severe physical or cognitive impairments face challenges operating joystick-controlled electric wheelchairs. Human-following smart wheelchairs provide a more accessible alternative, yet deploying deep learning-based detection on embedded devices is often constrained by computational demands. This study proposes an ID-based human tracking system that combines a modified YOLOv8n with C2fGhost modules for lightweight detection and DeepSORT for robust identity tracking. The proposed model achieves accuracy comparable to the baseline YOLOv8n with mAP50 of 0.624 compared to 0.628 while reducing parameters by 7.9% and computational load by 4.9%. In low-light scenarios, it improves detection accuracy with MOTA of 98.97% compared to 89.66% for the baseline, although identity consistency decreases with IDF1 of 67.24% compared to 94.83%. These findings demonstrate that the proposed approach maintains competitive accuracy with lower computational cost, making it suitable for embedded deployment not only in smart wheelchair navigation but also in broader edge AI applications requiring real-time and resource-efficient tracking. © COPYRIGHT SPIE. Downloading of the abstract is permitted for personal use only.
Department of Informatics Engineering, Faculty of Computer Science, Brawijaya University, Malang, Indonesia
