I have to design a system on a mobile platform which can be use to detect humans. In this case I'm going to use a PTZ high-definition camera.
I'm totally new to this so where should I start writing an OpenCV program and are there any papers which could help me?
Hello.
I'm agree with Marek Kraft that HOG (Dalal Triggs) is most common way to detect pedestrian. But there is another ways:
1. You can also use cascade detection (Viola Jones) on LBP it is faster then HOG, but less accuracy. There is some tricks to increase accuracy with cascades: use many different representation of image like edges or gradient.
2. Another way of HOG is LatentSvmDetector. It has more accuracy then HOG but slower, because it uses many detectors and recalculate some features on higher res (HOG with cells 4x4 instead standard 8x8).
3. Interesting work out of OpenCV "Pedestrian detection at 100 frames per second"http://www.cs.huji.ac.il/~peleg/CVPR2012/data/papers/367_O3B-04.pdf . With nice trick (like octave in SIFT) that uses pyramid with integer values 2, 4.... and subresolution covers with different detectors.
Hello,
The tutorial here:
http://chrisjmccormick.wordpress.com/2013/05/09/hog-person-detector-tutorial/
explains the basic concepts behind the probably most widely used human silhouette detection method and contains links to example OpenCV code.
Hello.
I'm agree with Marek Kraft that HOG (Dalal Triggs) is most common way to detect pedestrian. But there is another ways:
1. You can also use cascade detection (Viola Jones) on LBP it is faster then HOG, but less accuracy. There is some tricks to increase accuracy with cascades: use many different representation of image like edges or gradient.
2. Another way of HOG is LatentSvmDetector. It has more accuracy then HOG but slower, because it uses many detectors and recalculate some features on higher res (HOG with cells 4x4 instead standard 8x8).
3. Interesting work out of OpenCV "Pedestrian detection at 100 frames per second"http://www.cs.huji.ac.il/~peleg/CVPR2012/data/papers/367_O3B-04.pdf . With nice trick (like octave in SIFT) that uses pyramid with integer values 2, 4.... and subresolution covers with different detectors.
This is so much of helpful. Thank you so much Marek Kraft and Nikolay Sergievskiy. highly appreciated.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing