Automated Human Physical Function Measurement Using Constrained High

Automated Human Physical Function Measurement Using Constrained High Dispersal Network With SVM-Linear Dan Meng 1, Guitao Cao 1, *, Xinyu Song 1, Weiting Chen 1, Wenming Cao 2, + 1 East China Normal University 2 Shenzhen University *Corresponding Author Email: gtcao@sei. ecnu. edu. cn +Co-corresponding Author Email: wmcao@szu. edu. cn

1. Introduction • Critical problem in pattern recognition and computer vision – To discover suitable representations of images. – Rare reports on physical function measurement methods. – Physical function measurement can be considered as a human action classification problem with fine-grained.

1. Introduction • Why feature representations are important? – The performance is heavily dependent on the choice of features. • Solution – Hand-crafted features (SIFT, HOG), deep learning features.

1. 1 Motivations • Goal – Find the trade-off between complexity and performance. – Learn useful features from the clinical data in an unsupervised way. – Learn how to partition a human’s physical function into normal or abnormal state. • How? – Learn features from global data. – Use the most basic and easy operations.

1. 2 Contributions • A nonlinear transformation layer after 2 DPCA convolution. • Multi-scale feature pooling layer – Multi-scale histogram. – High dispersial. – Local response normalization. • Extremely simple and efficient

2. System Overview

2. Methodology 2. 1 Database Creation Collected by Kinect: 3840 physical depth images, with 12 different physical statuses, and each status was performed by 16 unique subjects for 20 times.

2. Methodology 2. 2 Feature Extraction 2. 2. 1 Feature Convolution Layer 2. 2. 2 Non-linear Transformation Layer 2. 2. 3 Multi-scale Feature Pooling Layer 2. 2. 4 Classifier

2. 2. 1 Feature Convolution Layer

2. 2. 1 Feature Convolution Layer

2. 2. 1 Feature Convolution Layer

2. 2. 3 Multi-scale Feature Pooling Layer

2. 2. 4 Classifier The compared classification methods include: • Support Vector Machine With Linear kernel(SVMLinear) • Support Vector Machine with RBF Kernel(SVM-RBF) • Naive Bayes(NB) • K Nearest Neighbors(KNN) • Random Forest(RF) • Decision Tree Using CART(DT) • Gradient Boosting Decision Tree(GBDT).

3 Results and Discussion • Impact of Feature Dimension Accuracy of different feature dimensions Time spend of different feature dimensions

3 Results and Discussion • Impact of Classifiers • Impact of the number of training samples

3 Results and Discussion • Impact of Feature Extraction Methods Accuracy of different feature extraction methods Confusion matrix of proposed CHDNet

Thank you!