Fitri Utaminingrum, Aldiansyah Satrio Kabisat, Anindya Zulva Larasati, Aulia Riza Mufita, Fais Al Huda, Achmad Arwan
Wheelchairs are vital mobility aids, yet existing control interfaces remain inadequate for users with severe disabilities who cannot operate them independently. Visual tracking-based human-following systems offer a promising solution, but correlation filter-based trackers (CFTs) have not been systematically evaluated in this context. This study compares four representative CFT algorithms - Minimum Output Sum of Squared Error (MOSSE), Kernelized Correlation Filter (KCF), Channel and Spatial Reliability Tracking (CSRT), and Efficient Convolution Operator (ECO) - for human-guide tracking in smart wheelchair navigation. Using 20 annotated video sequences comprising 24,648 frames under diverse conditions such as appearance variation, low light, occlusion, and similar-object interference, performance was assessed using precision, normalized precision, and success rate. Results show that ECO achieved the best overall performance with precision 0.4328, normalized precision 0.7645, and success rate 0.7302 at 30 FPS, outperforming CSRT, KCF, and MOSSE. These findings imply that ECO-based human-following systems can enhance safety, reduce caregiver burden, and support practical deployment in assistive mobility technologies. © COPYRIGHT SPIE. Downloading of the abstract is permitted for personal use only.
Department of Informatics Engineering, Faculty of Computer Science, Brawijaya University, Malang, Indonesia
