Semantic Object Detection for Human Activity Monitoring System

Abstract

Semantic object detection is significant for activity monitoring system. Any abnormalities occurred in a monitored area can be detected by applying semantic object detection that determines any displaced objects in the monitored area. Many approaches are being made nowadays towards better semantic object detection methods, but the approaches are either resource consuming such as using sensors that are costly or restricted to certain scenarios and background only. We assume that the scale structures and velocity can be estimated to define a different state of activity. This project proposes Histogram of Oriented Gradient (HOG) technique to extract feature points of semantic objects in the monitored area while Histogram of Oriented Optical Flow (HOOF) technique is used to annotate the current state of the semantic object that having human-andobject interaction. Both passive and active objects are extracted using HOG, and HOOF descriptor indicate the time series status of the spatial and orientation of the semantic object. Support Vector Machine technique uses the predictors to train and test the input video and classify the processed dataset to its respective activity class. We evaluate our approach to recognise human actions in several scenarios and achieve 89% accuracy with 11.3% error rate.