Record Details

<p>A Feature Weighting Technique on SVM for Human Action Recognition</p>

Online Publishing @ NISCAIR

View Archive Info
 
 
Field Value
 
Authentication Code dc
 
Title Statement <p>A Feature Weighting Technique on SVM for Human Action Recognition</p>
 
Added Entry - Uncontrolled Name Mishra, Soumya Ranjan; Department of Computer Science and Engineering, NIT Durgapur
Krishna, K Deepthi; Department of Computer Science and Engineering, VMTW, JNTUH
Sanyal, Goutam ; Department of Computer Science and Engineering, NIT Durgapur
Sarkar, Anirban ; Department of Computer Science and Engineering, NIT Durgapur
 
Uncontrolled Index Term Histogram of gradient; Interest point; Optical flow; Weighted feature SVM
 
Summary, etc. <p class="Abstract">Human action recognition is a challenging research topic and attracted very good attention in the last few years. This paper presents a features weighting framework for human action recognition based on the movement of different body-parts. Intuitively, Understanding the motion of a particular body-part having a major contribution to a specific action gives a better representation of that human activity. For example, action like walking, running and jogging, movement of the leg is more important and in boxing, waving and clapping, hand movement is more effective. This work presents a technique, utilizing the sub-region body-parts recognition rate to the weight kernel function. First, the complete human body is extracted from the background and HOG (histogram of gradient) based body-part detection is applied to generate three different sub-region (head, arm and body, foot and leg) of complete human body. Recognition rate and weight is calculated for all these sub-region (body-parts) for a particular action. Based on the weight (<em>ω</em>) of sub-region, a weighted feature Gaussian kernel function is obtained and weighted feature support vector machine (WF-SVM) classifier is constructed. The experimental results of the proposed framework have better performance on both KTH and UCF-ARG datasets compared against several state-of-the-art methods.</p>
 
Publication, Distribution, Etc. Journal of Scientific and Industrial Research (JSIR)
2020-11-09 16:10:07
 
Electronic Location and Access application/pdf
http://op.niscair.res.in/index.php/JSIR/article/view/40477
 
Data Source Entry Journal of Scientific and Industrial Research (JSIR); ##issue.vol## 79, ##issue.no## 7
 
Language Note en