Feature selection for grasp classification Lillian Chang November 29 2006 10 701 project final report Abstract Although the human hand is a complex biomechanical system functional grasps may be described by small set of features Supervised feature selection is used to evaluate the performance of reduced marker sets for grasp classification from motion capture data Our reduced feature set maintains 85 grasp classification accuracy compared to 90 accuracy from using the full 30 marker set Using a linear classifier and as few as 5 surface markers allows for dramatic simplification of the experimental procedure and reduced computational cost for grasp classification 1 Introduction The human hand has several degrees of freedom which provide it amazing flexibility as a manipulator but pose challenges for measuring and modeling hand movement One technique for measuring the human motion is motion capture where optical markers attached to body segments are used to reconstruct joint movement Motion capture of the hands is difficult because of marker occlusions due to the wide range of hand poses and close position of the fingers A reduced marker set would simplify the capture procedure and also describe hand configuration in a low dimensional space Furthermore there is no standardized motion capture protocol for the hand The aim of this study is to evaluate how the selected marker positions on the hand can affect the measured motion and investigate which markers may be the best to include in a reduced hand marker protocol for future experiments 2 Related work Previous literature in biomechanics has also investigated how to represent grasping and reach to grasping actions in a low dimensional feature space using principal components analysis For example 1 2 suggest that hand grasping motions can be represented by just a few principal components in the joint space These studies used a data glove to capture the motion of the hand and the experiments involved mimed hand motion or reach to grasp motion before the hand was in contact with the object Other work in the robotics community has investigated grasp classification One study 3 used neural networks for predicting grasps based on a taxonomy proposed in 4 The experiments focused on only three object shapes and sizes Hidden Markov Models have also been used for grasp recognition from data glove measurements 5 Data gloves can be cumbersome to the user may affect the natural grasping motion and do not always fit individual subjects well In our study we focus on input data in the context of marker based methods for motion capture and examine how to design an appropriate protocol which can simplify the data acquisition procedure 3 Problem definition This study investigates feature selection of the marker position inputs in conjunction with grasp classification The classification goal is to predict the grasp at a single time frame given the measured marker positions representing the hand configuration The purpose of feature selection is to evaluate which markers are the best predictors of the grasp class and determine which markers could be eliminated to simplify the marker protocol We aim to explore questions such as What is the minimum number of markers needed to represent the hand pose well and where on the hand should those markers be placed The redundancy in hand grasping motion found by 1 2 suggests that the number of markers could be dramatically reduced without severely compromising grasp classification 4 4 1 Proposed method Grasp classification Our approach uses linear classifiers for predicting grasp which will then be combined with supervised feature selection Although the fingers do exhibit nonlinear kinematics relative to the palm the constraints on hand motion will limit each surface marker to a small set of clustered reachable positions Different grasps are characterized by the relationships between marker positions which will vary but not in a severely nonlinear way Thus we believe that a classifier with linear decision boundaries can be successful for predicting grasp types from motion capture data In addition using linear classifiers can provide a simpler implementation and reduced computational cost compared to neural networks and Hidden Markov Models as used in 3 5 We will evaluate three candidates for a baseline classifier which uses the full feature set for predicting grasps Gaussian Naive Bayes GNB with class independent variances multiclass logistic regression LR and linear support vector machines SVM We expect that GNB will be less successful than LR and SVM due to the assumptions of conditional independence and Gaussian distribution of the marker coordinates which are unlikely to be satisfied in our case of coordinated hand movement 4 2 Supervised feature selection Given a single baseline classifier we wish to select a subset of features which simplifies the model We use two standard approaches for supervised feature selection as described in 6 First in the filter approach single features are individually ranked by a scoring criterion and the reduced set consists of the k best scoring features The advantage of this method is that each feature need only be scored once with the expense that the final selection does not consider possible interdependencies between the features We evaluate two scoring criteria 1 the mutual information between the target value and a single feature and 2 the prediction accuracy of a single feature classifier on a validation test set To compute the mutual information with the discrete target value the continuous marker position features are discretized rather than fit to an assumed distribution Wrapper methods are a second approach for feature selection In contrast to the filter approach wrapper algorithms consider the interaction between features in constructing the reduced set By modeling how the set of features is related to the target attribute rather than only the relation between each single feature with the target wrapper methods can potentially select a feature set of the same size which results in better prediction However this requires additional computational cost for training as the features must be re scored each time the current feature set changes To avoid considering the exponential number of possible feature sets the simplest wrapper algorithms consider a single feature at a time for locally optimal feature selection In this work we will consider two versions of greedy
View Full Document
Unlocking...