3D Gait Recognition Using Spatio-Temporal Motion Descriptors

Abstract

Abstract. We present a view independent algorithm for 3D human gait recognition. The identification of the person is achieved using motion data obtained by our markerless 3D motion tracking algorithm. We report its tracking accuracy using ground-truth data obtained by a markerbased motion capture system. The classification is done using SVM built on the proposed spatio-temporal motion descriptors. The identification performance was determined using 230 gait cycles performed by 22 persons. The correctly classified ratio achieved by SVM is equal to 93.5% for rank 1 and 99.6% for rank 3. We show that the recognition performance obtained with the spatio-temporal gait signatures is better in comparison to accuracy obtained with tensorial gait data and reduced by MPCA.

Human gait recognition from motion capture data in signature poses

“…Figure 6 shows that our method achieves the best Method EER AUC MAP Ahmed [9] 0.2779 0.7811 0.7948 Andersson [10] 0.2858 0.7755 0.7996 Balazia 0.1674 0.8902 0.8907 Ball [11] 0.3949 0.6714 0.6931 Dikovski [13] 0.2273 0.8288 0.8287 Gavrilova [14] 0.1922 0.8521 0.8904 Jiang [15] 0.2393 0.8414 0.8497 Krzeszowski [16] 0.2287 0.8464 0.8593 Kumar [17] 0.3545 0.6528 0.6197 Kwolek [18] 0.1839 0.8796 0.8840 Preis [19] 0.4916 0.5225 0.5356 Sedmidubsky [20] 0.2823 0.7726 0.7558 Sinha [21] 0.2194 0.8289 0.8493 Table 4 EER, AUC and MAP derived from FAR/FRR, ROC and recall/precision, respectively, plotted in Figure 6.…”

“…The databases of all surveyed research groups are either acquired by Kinect tracking a small area [9,13,16,17,18,19] or contain pre-calibrated skeletons of all participants [20]. Another problem is their size: The largest structural gait database currently available has 230 gait cycles of 22 walking subjects.…”

“…To describe characteristic aspects of gait pattern, various kinds of motion features can be extracted at the level of individual frames from the joint coordinates, such as relational features [23], joint angles [10,11,13,14,16,18], inter-joint distances [13,14,16,18,20,21], velocities or acceleration. Over a human body model (of any underlying joint structure) we denote descriptor as a set D = {F 1 , .…”

An Evaluation Framework and Database for MoCap-Based Gait Recognition Methods

“…They are typically combinations of static body parameters (bone lengths, person's height) [18] with dynamic gait features such as step length, walking speed, joint angles and inter-joint distances [3,6,18,20], along with various statistics (mean, standard deviation or maximum) of their signals [5]. Clearly, these features are schematic and human-interpretable, which is convenient for visualizations and for intuitive understanding, but unnecessary for automatic gait recognition.…”

“…Gait samples are compared by a distance function of their covariance matrices. • Kwolek by Kwolek et al [18] processes signals of bone angles around all axes, the person's height and step length. The gait cycles are normalized to 30 frames.…”