Combining Shape of Trajectories with MHI and their Directional Derivative-Based Description for Human Activity Recognition

This research introduces a unified framework for human activity recognition that integrates global temporal characteristics, local spatial information, and trajectory shape cues. Trajectory shapes are extracted by tracking key points using a Motion History Image (MHI) as a mask, eliminating the need for unreliable key-point and trajectory tracking. The selected key points from both the intensity image (local spatial information) and the MHI (global temporal information) are represented using the Histogram of Directional Derivative (HODD) descriptor, which effectively captures their visual and structural attributes. The combined feature representation is encoded through a Bag-of-Visual-Words (BoVW) model, and classification is performed using a multiclass Support Vector Machine (SVM). Extensive experiments on four benchmark datasets—URADL, KTH, Weizmann, and UCF101—yield accuracies of 95.4%, 95.83%, 100%, and 89%, respectively, demonstrating robustness to illumination changes, occlusion, and background clutter, and outperforming several state-of-the-art methods. Overall, the proposed framework offers a computationally efficient and highly discriminative solution for human activity recognition by effectively fusing trajectory shape, spatial, and temporal descriptors.