| dc.description.abstract | This paper aims to identify the major object detection techniques performed on assistive
technologies for visually impaired individuals, with a particular focus on the shift from
conventional computer vision methodologies to current deep learning frameworks.
The most popular object detection models under discussion are single-stage detectors:
Single Shot Detector (SSD) and You Only Look Once (YOLO), and two-stage detectors:
Faster R-CNN and RetinaNet, all evaluated in terms of performance in per-pixel
accuracy on the COCO dataset, expressed through mean Average Precision (mAP)
and inference time on GPU. The proposed models are discussed in the context of
their practical applicability for assistive purposes by considering problems such as
small object recognition, time constraints, fluctuations in the environment, energy
demands, limitations of devices, and interfaces. Observations show that single-stage
detectors such as the SSD and YOLO can provide faster inference time, which is ideal
for real-time application at 22ms and 29ms, though with lower mAP performance than
two-stage detectors in research with 23.2 and 33.0, respectively. The two-stage detectors
such as Faster R-CNN and RetinaNet are more accurate with 36.2% and 37.8% mAP
respectively, but have higher inference times of 200ms and 73ms for real-time assistive
tasks. The study also highlights the issues in small and occluded object detection, that
can be detected in diverse lighting and weather conditions, and power and hardware
constraints of wearable technology. These problems must be addressed by optimizing
the model and enhancing the hardware and software of assistive technologies for the
visually impaired. | en_US |
