Abnormal Human Action Recognition using Average Energy Images

In present scenario, Abnormal Human Activity Recognition (AbHAR) system has potential applications in smart homes, Ambient Assistive Living (AAL) and healthcare services. In this paper a novel Average Energy Image (AEI) based feature descriptor is designed for abnormal human activity recognition by integrating HOG and PCA with AEI. AEI helps to encode all the body shape variations for one activity in one frame effectively. Depth frames are used to generate binary silhouettes which are summed up to form an AEI per activity. AEI holds spatiotemporal shape variations which are encrypted as a feature vector using Histogram of Oriented Gradients (HOG) followed by Principal Component Analysis (PCA). The proposed approach is validated on publically available UR Fall Detection Dataset and newly introduced Abnormal Human Activity (AbHA) Dataset which delivers 92.5% and 98.9% accuracy.