High Dimensional Abnormal Human Activity Recognition Using Histogram Oriented Gradients and Zernike Moments

In this paper a robust abnormal human activity recognition framework is proposed which goals to recognize any unusual activity to the elderly people and to strengthen the concept of independent and quality living. The framework is structured to construct a robust feature vector by computing integrated feature vector: Histogram of Oriented Gradients (HOG) and Zernike moments on Average Energy Images (AEI). Formation of AEI images provides a compact representation of the video sequences without any Spatio-temporal loss of information. Integration of HOG and Zernike moments augments inter-class separation and translational and rotational invariance. The depth silhouettes are acquired by Microsoft's Kinect sensor which are used to generate clean binary silhouettes by background subtraction making the pre-processing faster and simpler with accuracy. The combined feature vector dimensions are reduced by applying PCA and SVM is applied to classify the activities. The proposed work is validated on publically available UR fall detection and Kinect Activity Recognition Dataset (KARD) 3D dataset. The experiments exhibit impressive results with The average recognition accuracy achieved on these datasets are 94% and 95.22% ARA for UR fall dataset, and KARD dataset, respectively.