Robot service for elderly to find misplaced items: A resource efficient implementation on low-computational device

Elderly people often forget to put the items they need due to decreased memory. In this study, we developed an Integrated platform assistance robot providing support to elderly people. We developed a robot assistant platform that was equipped with an indoor positioning system that can help the elderly find misplaced items. Deep learning already has good accuracy in detecting the object but requires great computation resources. When applied to devices that have limited computing and memory capabilities such as robots, the computation time becomes slow or not applicable. We built a lightweight CNN that could run on a single board computer. To improve the accuracy of the network, we apply knowledge distillation by using an extensive network (YOLOv3) as a teacher. To increase computational speed, we do it by reducing the number of layers by implementing batch normalization fission. After being tested on the YOLO, knowledge distillation method can be used to increase accuracy, batch normalization fission will increase computation speed. From the experiment results using the VOC dataset on YOLO architecture with MobileNet feature extractor, the knowledge distillation method can increase accuracy by 9.4% from 0.3850 mAP to 0.4215 mAP and batch normalization fission can speeds up the computation time to 100.7% from 8.3 FPS to 16.66 FPS on CPU i7. The Knowledge Distillation successfully increase the model's accuracy, reducing the model's size, and batch normalization fusion method can speed up the detection process.