Feature extraction for human action recognition based on saliency map
Human Action Recognition (HAR) plays an important role in computer vision for the interaction between human and environments which has been widely used in many applications. The focus of the research in recent years is the reliability of the feature extraction to achieve high performance with the us...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2018
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/79551/1/TanYiPingMFKE2018.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Human Action Recognition (HAR) plays an important role in computer vision for the interaction between human and environments which has been widely used in many applications. The focus of the research in recent years is the reliability of the feature extraction to achieve high performance with the usage of saliency map. However, this task is challenging where problems are faced during human action detection when most of videos are taken with cluttered background scenery and increasing the difficulties to detect or recognize the human action accurately due to merging effects and different level of interest. In this project, the main objective is to design a model that utilizes feature extraction with optical flow method and edge detector. Besides, the accuracy of the saliency map generation is needed to improve with the feature extracted to recognize various human actions. For feature extraction, motion and edge features are proposed as two spatial-temporal cues that using edge detector and Motion Boundary Histogram (MBH) descriptor respectively. Both of them are able to describe the pixels with gradients and other vector components. In addition, the features extracted are implemented into saliency computation using Spectral Residual (SR) method to represent the Fourier transform of vectors to log spectrum and eliminating excessive noises with filtering and data compressing. Computation of the saliency map after obtaining the remaining salient regions are combined to form a final saliency map. Simulation result and data analysis is done with benchmark datasets of human actions using Matlab implementation. The expectation for proposed methodology is to achieve the state-of-art result in recognizing the human actions. |
|---|