Multi-view camera modules for human action recognition
With the advancements of technology, human action recognition is still one of the recurrent problems in computer vision. A multi-view modular camera system to identify human actions in an indoor environment is proposed in this paper to carry off the viewpoint problems. Histogram of Oriented Gradient...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | text |
| Language: | English |
| Published: |
Animo Repository
2016
|
| Online Access: | https://animorepository.dlsu.edu.ph/etd_bachelors/10988 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | De La Salle University |
| Language: | English |
| Summary: | With the advancements of technology, human action recognition is still one of the recurrent problems in computer vision. A multi-view modular camera system to identify human actions in an indoor environment is proposed in this paper to carry off the viewpoint problems. Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) was used for person detection. Then, timed Motion History Images (tMHI) was used to represent an action and the extracted features was used for Hidden Markov Model (HMM) to classify actions. Afterwards, a decision module was used to sync and combine the results of all cameras. Furthermore, the system went different tests using our own recorded dataset that is composed of known and unknown actions. Considering only the known actions, the system has an accuracy rate of 83.33% while including the unknown actions resulted into an accuracy rate of 43.52%. It was found that similar actions are the main factor that usually result into misclassification due to lack of discriminatory influence of the feature descriptor. Other factors included are the noise reduction method used and the camera's viewpoint. Overall, the system showed promising results as the aim was to create a practical and deployable solution to the main problem. |
|---|