Daily Living Activities Recognition via Efficient High and Low Level Cues Combination and Fisher Kernel Representation

Abstract

In this work we propose an efficient method for activity recognition in a daily living scenario. At feature level, we propose a method to extract and combine low-and high-level information and we show that the performance of body pose estimation (and consequently of activity recognition) can be significantly improved. Particularly, we propose an approach extending the pictorial deformable models for the body pose estimation from the state-of-the-art. We show that including low level cues (e.g. optical flow and foreground) together with an off-the-shelf body part detector allows reaching better performance without the need to retrain the detectors. Finally, we apply the Fisher Kernel representation that takes the temporal variation into account and we show that we outperform state-of-the-art methods on a public dataset with daily living activities.

A modified vector of locally aggregated descriptors approach for fast video classification

“…They detect small bounding boxes with a multi-scale HoG descriptor, instead of complete body limbs, making their work more efficient because it prevents the problem of double counting. The body part detector combined with the HoF features obtained good results on daily living activities [49]. However, these framework are adapted to a specific task and requires the use of motion compensation for foreground estimation and the detection and tracking of the human in the scene, generating a high computational cost.…”

“…Human pose and body-part motion obtained good results in many event detection categories [48,49,66]. We extract the body-part components using the state-of-the-art body-part detector [48] and compute at every frame for all 18 body-parts a Histogram of optical Flow in 8 orientations [49];…”

“…We did not use audio and text information because the movies from UCF101 dataset do not contain this information. For the ADL dataset we use only the body-part features which already provided state-of-the-art results in many approaches [49].…”

“…We use optical flow at each pixel obtained using the Lucas-Kanade method [29] and apply a threshold on the magnitude of the optical flow, to decide if the pixel is moving or is stationary. We divide the frames in 3x3 regions and then we compute the HoF feature for each region [32]; -Body-part features (144 values) [49] approximate the optical flow that is computed on the body-part components. Human pose and body-part motion obtained good results in many event detection categories [48,49,66].…”

“…We divide the frames in 3x3 regions and then we compute the HoF feature for each region [32]; -Body-part features (144 values) [49] approximate the optical flow that is computed on the body-part components. Human pose and body-part motion obtained good results in many event detection categories [48,49,66]. We extract the body-part components using the state-of-the-art body-part detector [48] and compute at every frame for all 18 body-parts a Histogram of optical Flow in 8 orientations [49];…”