Using Multi Class SVM for Creating an Algorithm for Human Activity Recognition

Identifying human activity is a crucial task in numerous fields involving human-computer interaction. This study introduces a innovative approach for human activity recognition using a multi-class Support Vector Machine (SVM). The methodology unfolds in three crucial steps: preprocessing, feature extraction, and classification. In the preprocessing stage, a depth map undergoes transformation into a silhouette for each frame, achieved by segregating foreground elements from the background. To mitigate the effects of variations in size and position, foreground silhouettes undergo normalization via shrinking and cropping. The process of feature extraction includes deriving spatial and temporal features from these normalized silhouettes. This involves integrating the directed gradients histogram, optical flow histogram, the motion history image, and the 2D Discrete Cosine Transform. Each frame is represented by a comprehensive feature vector obtained by concatenating these extracted features. These aggregated features are subsequently utilized for activity identification using a multi-class SVM classifier. Experimental results on three benchmark datasets validate the effectiveness of this approach, achieving commendable accuracy rates in activity recognition. A notable strength of the proposed method is its robustness to occlusions, changes in illumination, and variations in view angles, making it a compelling choice for accurate and reliable human activity identification in real- world scenarios.