Gait Recognition Using HMMs and Dual Discriminative Observations for Sub-Dynamics Analysis

TLDR: A new gait recognition method that combines holistic and model-based features that is able to capture more detailed sub-dynamics by refining upon the preceding general dynamics.

Abstract: We propose a new gait recognition method that combines holistic and model-based features. Both types of features are extracted automatically from gait silhouette sequences and their combination takes place by means of a pair of hidden Markov models. In the proposed system, the holistic features are initially used for capturing general gait dynamics whereas, subsequently, the model-based features are deployed for capturing more detailed sub-dynamics by refining upon the preceding general dynamics. Furthermore, the holistic and model-based features are suitably processed in order to improve the discriminatory capacity of the final system. The experimental results show that the proposed method exhibits performance advantages in comparison with popular existing methods.

Figures

Fig. 5. The two HMMs used for gait recognition (four states instead of the actual five are shown). (a) The dominant HMM, which represents general shape and dynamics of gait using holistic features (observations). All states generate observations. (b) The refinement Hierarchical HMM that represents shape details and sub-dynamics of gait using detailed labelled component features. Note that only states at the lower level of the refinement HMM generate observations. The states labelled “E” in the lower level of the refinement HMM are “end” pseudo-states and do not generate observations. The “end” states indicate a return to the upper level and a transition to the next upper-level state. For this reason, transition to an “end” state (indicated using dotted arrows) practically means a transition to the first lower-level state of the next HMM.

Fig. 8. Cumulative match scores for gait recognition. Top row: system based on holistic features only. (a) original features, (b) discriminative features (HL method). Bottom row: system based on the combination of holistic and labelled component features for analysis of sub-dynamics. (c) original features, (d) discriminative features (DHM method).

TABLE IV RECOGNITION RATES ACHIEVED BY THE PROPOSED METHOD USING SIMILARITY MEASURES BASED ON EQUATIONS (16), (17),(21), AND (22).

TABLE I STEPS FOR THE LABELLING OF BODY COMPONENTS

Fig. 1. Block diagram of the proposed algorithm: (a) training, (b) testing.

TABLE III EIGHT DIFFERENT SETS IN THE ACTIBIO DATABASE.

Frequently asked questions

Q1. What have the authors contributed in "Gait recognition using hmms and dual discriminative observations for sub-dynamics analysis" ?

The authors propose a new gait recognition method that combines holistic and model-based features. Furthermore, holistic and model-based features are suitably processed in order to improve the discriminatory capacity of the final system. 

Q2. What is the probability of an observation being generated by a left-to-right HMM?

Since the order with which poses are assumed during walking is defined naturally, a left-to-right HMM is considered, the initial state probabilities of which areπn = 1 if n = 10 otherwise (5) where πn is the initial state probability for the nth state, n = 1, . . . , N , of the dominant HMM. 

Q3. Why is it possible to calculate the gravity centre of the lth body component?

Due to the availability of labelling information for each pixel, it is possible to calculate the gravity centre gl of the lth body component. 

Q4. What is the rationale for using a refinement HMM?

The rationale for using a refinement HMM is that it can use the detailed labelled body component features in order to refine the general shape/dynamics that were captured previously through the dominant HMM. 

Q5. What is the performing version of their proposed system?

The best performing version (DHM) of their proposed system improves on HL, providing further evidence that the additional deployment of labelled component features can improve the performance of holistic systems by analyzing gait sub-dynamics. 

Q6. How do the authors segment a gait cycle out of a silhouette?

the authors segment a gait cycle out of a gait silhouette sequence using the method proposed in [22], i.e., by constructing a signal representing the number of pixels in each silhouette, filtering the signal by taking into account its autocorrelation, and locating the “peaks” in the filtered signal. 

Q7. What are the parameters that are to be determined for the lower-level HMMs?

The parameter calculation for the refinement lower-level HMMs is based on the observations F. Since the initial state probabilities are as in eq. (7), the parameters that are to be determined are the state-transition probabilities as well as the exemplars for each of the HMMs at the lower level of the refinement HMM. 

Q8. What is the comprehensive approach for reconstructing body components?

The Layered Deformable Model (LDM) introduced in [12] is the most comprehensive approach for reconstructing body components based on both automatically-segmented silhouettes and manually-labelled silhouettes. 

Q9. What is the probability that the observations are generated by the dominant HMM model?

following the derivations in [20], the probability that observations H are generated by the dominant HMM model λ isP (H|λ) = ∑ allQ P (H,Q|λ) (15)where Q is a sequence of states in the dominant HMM that can generate the observations. 

Q10. What is the discriminative transform for the exemplars?

But since the exemplars represent specific walking poses, theMay 30, 2013 DRAFT22discriminative transform derived from these exemplars indicate the discriminative differences in that specific pose.