Abstract
Human activity recognition system is an essential requirement for a service robot to understand its environment and avoid obstacles. However, the robot safety feature requires the hazard level of the detected object to decide on necessary action. This study aims to develop a model to identify the type of human activity and determine the hazard level for the robot by using a region of interest-based decision-making approach. A total of 1900 images of the five most potentially hazardous activities in the hospital environment from a robot perspective are collected and used for training. Three deep learning models, namely, YOLOv2, VGG16, and MobileNetv2 SSD, are used to classify hazardous activity. Experimental results using the Deeplearning4J (DL4J) and TensorFlow frameworks show that the VGG16 model exhibits the highest performance with an accuracy of 93.33%. The YOLOv2 and MobileNetv2 SSD models obtain an accuracy of 46.67% and 68.95%, respectively. The misclassification of activity in the hospital environment is due to the high similarity of activities. Further study should be performed by collecting more data in different classes in the actual hospital environments.