Cross View Gait Recognition Using Correlation Strength

Gait is a behavioural biometric particularly useful for non-intrusive and/or non-cooperative person identification from a distance in unconstrained public spaces. However, such environments also increase the difficulties in gait recognition compared to a more controlled one with constantly known view angle. This is largely because various factors can affect gait including people walking in different clothes, under different carrying conditions, at variable speed, in different shoes and from arbitrary views. In particular, changes in view angle pose one of the biggest challenges to gait recognition as it can change significantly the available visual features for matching. Recently a number of approaches [2, 3] based on view transformation have been presented which have the potential to cope with large view angle changes and do not rely on camera calibration. These approaches aim to learn a mapping relationship between gait features of the same subject observed across views. When matching gait sequences from different views, the gait features are mapped/reconstructed into the same view before a distance measure is computed for matching. An advantage of these methods is that they have better ability to cope with large view angle change compared to earlier works. However, a view transformation based method also has a number of drawbacks 1) it suffers from degeneracies and singularities caused by features visible in one view but not in the other when the view angle difference is large. 2) The reconstruction process propagates the noise present in the gait features in one view to another thus decreasing recognition performance. In this paper we propose a novel approach to cross view gait recognition by addressing the problems associated with the view transformation model. Specifically we model the correlation of gait sequences from different views using Canonical Correlation Analysis (CCA). A CCA model projects gait sequences from two views into two different subspaces such that they are maximally correlated. Similar to the existing view transformation methods, the CCA model also captures the mapping relationship between gait features of different views, albeit implicitly. However, rather than reconstructing gait features in the same view and matching them using a distance measure, we use the CCA correlation strengths directly to match two gait sequences. This brings out two key advantages: 1) by projecting the gait features into the two subspaces with maximal correlation, features that become invisible across views are automatically identified and removed. 2) without reconstruction in the original gait feature space, our approach is more robust against feature noise. In this paper we also address the problem of view angle recognition using Gaussian Process (GP) classification in order to build a complete gait recognition system with probe sequence view angle unknown. This differs from existing approaches which assume the probe view angle is known. Experiments are carried out to demonstrate that 1) our GP classification based view angle recognition method effectively identifies the view angle and is superior when compared to an SVM based method; 2) The gait recognition performance of our method significantly outperform those of the existing view transformation models [2, 3] even when they assume known probe sequence view angle. We compute two gait representations, one for view angle recognition and the other for cross-view gait recognition. For view recognition, gait sequences are represented using Truncated Gait Energy Images (TGEI). TGEI (Fig.1(bottom row))is simply Gait Energy Image (GEI) without its top part (head & torso) and is generated by only taking the bottom one third of the GEI. Gait Flow Images (GFI) [1] are used as a gait feature for cross view gait recognition. GFIs provide more discriminative representation for identity recognition compared to GEI by looking at multiple independent motion of different body parts during a gait cycle [1]. It is robust against various covariate conditions such as carrying and clothing [1]. (a) 36◦ (b) 54◦ (c) 72◦ (d) 90◦ (e) 108◦ (f) 126◦ (g) 144◦ Figure 1: Top row: GEIs of a same subject for different views. Bottom row: TGEIs obtained from the GEIs in the top row.