A data augmentation methodology for training machine/deep learning gait recognition algorithms

Abstract

In this paper we propose a data augmentation methodology for training machine/deep learning gait recognition algorithms. While previously published methods generated synthetic data to augment training and/or testing sets [1] or to learn features invariant to certain conditions [2], they have incorporated a very limited number of covariate factors. To our knowledge, this is the first attempt to provide the ability to simultaneously generate synthetic data with so many controllable conditions for gait recognition.The combination of real motion capture, data preparation, avatar construction and scripted rendering allows synthetic frames to be generated, with almost arbitrary degrees of variation. Figure 1 shows a small sample of the controllable confounding factors that can be generated with the proposed methodology. Using the Gait Energy Image (GEI) [3] as gait features, extracted directly from the synthetically generated frames, we performed experiments to assess the level to which identity is preserved within the synthetic data sets. Results from our experiments -presented in Figure 2 -suggest that information about the identity of subjects is retained within the synthetically generated data. The experiments using GEIs as features within our dataset showed that augmenting a limited amount of real data with the synthetically generated data can yield identification of subjects with an accuracy of more than 95%. The results of this study suggest that synthetic data can be used to augment the training of gait recognition algorithms provided that the feature set (e.g. principal components) is expanded to include both real and synthetic examples of data.

Hsc70-4 Deforms Membranes to Promote Synaptic Protein Turnover by Endosomal Microautophagy

“…In Jeon et al (2017), the authors studied augmentation of drone sounds using a publicly available dataset that contains several real-life environmental sounds. Furthermore, the research by Charalambous and Bharath (2016) explored the use of a DA method for training a deep learning algorithm for recognizing gaits. Another interesting use of data augmentation is the development of a model for 3D pose estimation using motion capture data (Rogez & Schmid, 2016).…”

AGA: Attribute-Guided Augmentation

“…While augmentation techniques can be as simple as flipping, rotating, adding noise, or extracting random crops from images [20,5,37], task-specific, or guided augmentation strategies [4,16,28,25] have the potential to generate more realistic synthetic samples. This is a particularly important issue, since performance of CNNs heavily relies on sufficient coverage of the variability that we expect in unseen testing data.…”

“…In [4], e.g., Charalambous and Bharath employ guidedaugmentation in the context of gait recognition. The authors suggest to simulate synthetic gait video data (obtained from avatars) with respect to various confounding factors (such as clothing, hair, etc.)…”

Time-Warping: A Time Series Data Augmentation of IMU Data for Construction Equipment Activity Identification

“…Automated, and real-time activity recognition of construction equipment plays an important role in construction operation analysis by enabling productivity monitoring [1][2], preparation of input for near real-time simulation [3][4][5], and automated cycle-time analysis [2][3][4][5][6]. It is also a key necessity for real-time safety applications on the construction site [7][8][9] and for automating environmental assessments [11][12]. Equipment activity identification can also enabled several applications in AR/VR visualization [12][13][14][15].…”