RPSA-YOLOv10: Relative Partial Self-Attention for Object Recognition in Smart Glasses Based on Contextual Adaptation

Uncertainty triggers everyone involved in activities in this world to adapt well, including visually impaired people.Therefore, this paper proposed a smart glasses model based on the Self-Adaptive Cyber-Physical System to help visually impaired people live their days.This model was equipped with object recognition capabilities as an extension of YOLOv10m and Relative Positional Encoding (RPE) where it was placed in the attention module and subsequently combined with Partial Self-Attention (PSA) to create a better understanding of spatial features compared to the attention module previously.Therefore, we introduced a new variant called Relative Partial Self-Attention (RPSA) YOLOv10.Our model indicated adaptability based on contextual knowledge, such as calculating object distances and the ability to work at low light intensity.Additionally, our model also operates through voice commands and voice notifications.The evaluative results of the model trained with the MS COCO dataset signified mAP50, mAP75, and mAP50-95 values of 67.3%, 54.5%, and 50.2% respectively with an inference speed of 8.4 ms/image.These results demonstrated better performance compared to other versions of the YOLO model, notably in evaluations using small datasets with an increase in mAP50-95 of 30.8% compared to the YOLOv10 model.In addition, our designed adaptive system can handle the problem of estimating object distances with an average error of 15.95%.Further, it can work on light intensity problems with a stable increase in average brightness reaching 95.65.