Human Activity Recognition Using Rotation Invariant Radial Fourier Profile Features

Human activity recognition (HAR) in video datasets is crucial for various applications, yet challenges persist in accurately identifying activities among diverse view and camera angles. Existing methods often struggle with rotational variations, leading to decreased recognition accuracy and performance gaps in real-world scenarios. To address this issue, this study proposes using Radial Fourier Profile (RFP) features extracted via 2D Fast Fourier Transform (FFT). RFP features inherently possess rotation invariance, offering a solution to the challenges posed by varying viewing conditions. The effectiveness of the proposed RFP feature extraction technique is evaluated using K-Nearest Neighbour (KNN) and Random Forest classifiers on UCF and HMDB datasets, representing diverse human activities and environmental conditions. Statistical tests, including ANOVA, are employed to assess the rotation invariance of RFP features. Results indicate significant improvements in activity recognition accuracy across various orientations, demonstrating the robustness of the proposed system.