I am working on traffic monitoring system,which i want to detect moving objects in a highway cameras and count number of each category of vehicles in video frames.
videos acquired by a static camera in an unchanged illumination condition.
I want to classify outputs into four category such as: bus and truck ,car , minibus and motorcycle.
i used optical flow for moving object detection, HoG features for modeling objects and then stacking HoG features in a vector and use distance measure between each input query from motion detection step and dataset which i make it before.
I am resizing all dataset and query images into equal size which is mean size of all images and then extract features.
I don't know is it right way of using HoG features in this way and also can anyone suggest any solution for different size of images in dataset and query image other than resizing all image.
I used HOG to detect human in disaster scene when I was a postgraduate student.
Our method is to extract HOG feature of our data base ,then sent them to a adaboost classifier to train the feature vectors. We us the classifier to detect human.
If you want to detect the vehicle by the distance of feature vector, I suggest that you can't take HOG as the feature, because HOG is a statistic feature that it can not describe the object in variation of global pattern. It means that the false alarm rate will very high.
I think the gray-level histogram is better when the measurement is vector distance.
You can consider the shape feature and ratio of the distance between angular points as the feature of the system. These features are invariant by the size of objects.
Shao, I was thinking about making features vectors of cars in various positions and use this vectors to train a classifier for each category.
I think your suggestion may have better result than HoG.Could you please give me more information about "shape feature and ratio of the distance between angular points as the feature of the system" ?
Ehsan, if you want to use vehicle-specific feature vectors then trained Haar cascades for different vehicular positions can be used. The issue is, you'd need a large training dataset for each of these positions to train such a classifier. Haar classifiers for object detection was proposed by Viola Jones (See link attached).
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.10.6807
Syed Yusuf, Many thanks for your good suggestion and solution.
Do you have any other suggestion for feature extraction method? Is HoG features good enough or may something else can be better than HoG?
You can use Code Book (CB) based background subtraction though that would under perform when the number of objects (vehicles) is really large. This technique however is very precise and computationally efficient if your scene contains no more than 10 - 15 objects at a time (See attached link). Alternatively, you can try Mixture of Gaussians (MOG). MOGs are predominantly worked-upon by Stauffer and Grimson (Link attached separately). However, this methodology fails to eliminate shadows or reflections in its most basic form. I personally believe you'd find it difficult to train a codebook CB classifier in a highly dynamic environment such as that of an active traffic scene. Yet, with a bit of pre-processing, MOGs can be used to obtain robust outlines of moving cars. You can also try blob-analysis. Another suggestion is to try getting an optical flow of your scene. Hope this all helps!!!
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5279646&tag=1
The MOG paper link is attached herewith. It seems you cannot add more than a single link in a post which is weird.
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=784637
Thank you for your clear and early response.Before i post this question, I have tested Optical flow and GMM to extract ROI of moving objects in video streams. I observed that in a clutter and occlusion condition,i cannot define exact region of individual objects and therefore i have some problem in these kind of situations.Also, in my research datasets, i don't have any clear background, and my method should be able to detect cars instantaneously.(But i have access to the dataset to make model of objects)
As a conclusion, i decided to use two stage detection. At first, Extract ROI using one of motion detection methods(like Optical flow and GMM), then perform some process to find cars,buses,minibuses and motorcycles from ROIs
Another choice, which i think about it is that i can find four categories without any pre-steps like motion detection and all the process could be done in feature extraction and data association steps.
Now, i am concentrating on a good feature extraction and classification method to be robust on occlusion and have low false detection rate.
I think, i should use fusion of some features and some dimension reduction steps to model object as good as i can.Finally, design classifier with acceptable response.
Ehsan,sorry to just read the inf. I have read some papers on shape features before,but I didn't take the relative experiments to test and verify. My future work will on the research of angular points. If I have some new ideas i will contact you.
Now, I can only send you some papers that I think is good. If you want to take some experiments on angular points, you can firstly set some angular points on the image manually. Then get the line between each two angular points. Crossratio of the corresponding lines is a projection invariant. So you can test if the crossratio can be a feature for object detection.
Many thanks for your detailed answers,Shao Chunyan. I read attached articles and the highlight parts. I have found Zernike Moment interesting, and tested it on matlab. Also,Generic Fourier Descriptor has acceptable results but i didn't find any implementation of this algorithm on net.
- Badges
- Science topic
- Similar topics
- Correction