Redefining histograms of oriented gradients descriptors for handling occlusion in people detection
Object detection is a big challenge for researchers to address the issues that affect accurate detections. The Histogram of oriented gradients (HOG) descriptors have been used extensively for object detection on challenging conditions with good results. However, occlusion remains a well-known issue...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2014
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/48597/1/AhmadSaniMFC2014.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Object detection is a big challenge for researchers to address the issues that affect accurate detections. The Histogram of oriented gradients (HOG) descriptors have been used extensively for object detection on challenging conditions with good results. However, occlusion remains a well-known issue where some parts of an object are only visible. Conventional frameworks using HOG descriptors for detecting whole human body are prone to occluded people and thus causing errors on implementations running these frameworks. This issue was addressed in different approaches like hybrid feature extraction algorithms. Rather than following such approaches, the problem is addressed in this research by redefining HOG descriptors to work in a part object (body) detection framework that is powered by a finite-state machine. Experiments were conducted to test a new method of extracting HOG descriptors for parts of human body via slicing a window’s HOG descriptor into four. The part body detection framework utilises each part’s descriptor whereas three detected parts are sufficient to declare that the window has an assumed instance. Support vector machines (SVM) were used to classify the extracted parts and a finite- state machine was employed for handling the detected parts. Training and testing were made on subsets from the INRIA Person Dataset and the results favoured the part body detection framework over the conventional whole body framework. For a test set of 50 positive images of occluded people, the part body detection framework successfully detected 46 true positive while the whole body framework detected 36. Moreover, the former had less false positive detection than the latter having 80 false positive windows comparing to 289. |
|---|