A Machine Learning Method with Threshold Based Parallel Feature Fusion and Feature Selection for Automated Gait Recognition

Gait﻿is﻿a﻿vital﻿biometric﻿process﻿for﻿human﻿identification﻿in﻿the﻿domain﻿of﻿machine﻿learning.﻿In﻿this﻿ article,﻿a﻿new﻿method﻿is﻿implemented﻿for﻿human﻿gait﻿recognition﻿based﻿on﻿accurate﻿segmentation﻿ and﻿multi-level﻿features﻿extraction.﻿Four﻿major﻿steps﻿are﻿performed﻿including:﻿a)﻿enhancement﻿of﻿ motion﻿region﻿in﻿frame﻿by﻿the﻿implementation﻿of﻿linear﻿transformation﻿with﻿HSI﻿color﻿space;﻿b)﻿ Region﻿of﻿Interest﻿(ROI)﻿detection﻿based﻿on﻿parallel﻿implementation﻿of﻿optical﻿flow﻿and﻿background﻿ subtraction;﻿c)﻿shape﻿and﻿geometric﻿features﻿extraction﻿and﻿parallel﻿fusion;﻿d)﻿Multi-class﻿support﻿ vector﻿machine﻿(MSVM)﻿utilization﻿for﻿recognition.﻿The﻿presented﻿approach﻿reduces﻿error﻿rate﻿and﻿ increases﻿the﻿CCR.﻿Extensive﻿experiments﻿are﻿done﻿on﻿three﻿data﻿sets﻿namely﻿CASIA-A,﻿CASIA-B﻿ and﻿CASIA-C﻿which﻿present﻿different﻿variations﻿in﻿clothing﻿and﻿carrying﻿conditions.﻿The﻿proposed﻿ method﻿achieved﻿maximum﻿recognition﻿results﻿of﻿98.6%﻿on﻿CASIA-A,﻿93.5%﻿on﻿CASIA-B﻿and﻿ 97.3%﻿on﻿CASIA-C,﻿respectively.


PROPOSEd METHOd
The proposed HGR approach comprises of four steps as shown in Figure 1.The preprocessing step improves human region in the given video sequence and later on extracts human region by implementinguniformsegmentationandoptimizingitseffectsbyfusionofbackgroundsubtraction.
Texture, geometric and shape features are extracted from ROI and parallel fusion is performed.Afterwards,afeaturesselectionalgorithmisimplementedtoselectthebestfeaturesfromfusedvector basedonminimumdistancethatarelaterrecognizedbyOneagainstallSVM.
The convolution filtering is used to modify the spatial properties of given frame because convolutionisageneral-purposefilterwhichaffectsgivenframeorimage.Anintegermatrixofsize 5×5isappliedonmotionframetodeterminethevalueofcentralpixelbythesumofallneighborsof weightedvalues.Finally,themeanvalueofresultantimageiscalculatedwhichconsistsofforeground andbackground.Themeanvalueisutilizedtosetathresholdfunctionforfinalbinaryimageas followsinEquation(13): The above equation shows that if difference between F i+1 and F i is less than mean value, backgroundisconsideredotherwiseforeground.Theeffectsofhumanextractionusingopticalflow andimprovedbackgroundsubtractionareshowninFigure3andFigure4.
Itmeansthatminimumdistancefeaturesarecombinedtogetherandmaximumdistancefeatures areremovedfromfinalselectedvector.Asaresult,anewminimumdistancevectorisobtainedwhich is later fed to MSVM for gait recognition.The cost function of MSVM is defined as: Where,

Figure 1 .
Figure 1.Flow diagram of proposed method

Figure 2 .
Figure 2. Frame preprocessing: a) original frame; b) HSI color transformation; c) Selection of hue channel; d) Weighted log transformation; e) Linear contrast stretching

Figure 3 .
Figure 3. ROI extraction using optical flow and improved background subtraction.a) Original frame; b) Horn-Schunck optical flow; c) Improved background subtraction; d) ROI extraction.
and ξ x 3 are three extracted feature vectors, where ξ x

Figure 5 .
Figure 5. Description of parallel based features fusion and selection

Figure 6 .
Figure 6.Visual results of proposed method for human gait recognition with label N (normal walk), W (wearing coat walk) and B (walk with bag)

Figure 7 .
Figure 7. Sample frames of selected data sets

Figure 10 .
Figure 10.Comparison of MSVM recognition accuracy with LDA, DT, and KNN on all three selected datasets