1 December 2014 in Computer Methods and Programs in Biomedicine
3 August 2019 in Multimedia Tools and Applications
We introduce synchronized and calibrated multi-view video and motion capture dataset for motion analysis and gait identification. The 3D gait dataset consists of 166 data sequences with 32 people. In 128 data sequences, each of 32 individuals was dressed in his/her clothes, in 24 data sequences, 6 of 32 performers changed clothes, and in 14 data sequences, 7 of the performers had a backpack on his/her back. In a single recording session, every performer walked from right to left, then from left to right, and afterwards on the diagonal from upper-right to bottom-left and from bottom-left to upper-right corner of a rectangular scene. We demonstrate that a baseline algorithm achieves promising results in a challenging scenario, in which gallery/training data were collected in walks perpendicular/facing to the cameras, whereas the probe/testing data were collected in diagonal walks. We compare performances of biometric gait recognition that were achieved on marker-less and marker-based 3D data. We present recognition performances, which were achieved by a convolutional neural network and classic classifiers operating on gait signatures obtained by multilinear principal component analysis. The availability of synchronized multi-view image sequences with 3D locations of body markers creates a number of possibilities for extraction of discriminative gait signatures. The gait data are available at http://bytom.pja.edu.pl/projekty/hm-gpjatk/.
12 May 2021 in Neural Computing and Applications
In this work, we present a new algorithm for human action recognition on raw depth maps. At the beginning, for each class we train a separate one-against-all convolutional neural network (CNN) to extract class-specific features representing person shape. Each class-specific, multivariate time-series is processed by a Siamese multichannel 1D CNN or a multichannel 1D CNN to determine features representing actions. Afterwards, for the nonzero pixels representing the person shape in each depth map we calculate statistical features. On multivariate time-series of such features we determine Dynamic Time Warping (DTW) features. They are determined on the basis of DTW distances between all training time-series. Finally, each class-specific feature vector is concatenated with the DTW feature vector. For each action category we train a multiclass classifier, which predicts probability distribution of class labels. From pool of such classifiers we select a number of classifiers such that an ensemble built on them achieves the best classification accuracy. Action recognition is performed by a soft voting ensemble that averages distributions calculated by such classifiers with the largest discriminative power. We demonstrate experimentally that on MSR-Action3D and UTD-MHAD datasets the proposed algorithm attains promising results and outperforms several state-of-the-art depth-based algorithms.
27 November 2017 in IET Computer Vision
The authors present efficient and effective algorithms for fall detection on the basis of sequences of depth maps and data from a wireless inertial sensor worn by a monitored person. A set of descriptors is discussed to permit distinguishing between accidental falls and activities of daily living. Experimental validation is carried out on the freely available dataset consisting of synchronised depth and accelerometric data. Extensive experiments are conducted in the scenario with a static camera facing the scene and an active camera observing the same scene from above. Several experiments consisting of person detection, tracking and fall detection in real‐time are carried out to show efficiency and reliability of the proposed solutions. The experimental results show that the developed algorithms for fall detection have high sensitivity and specificity.