How to handle occlusion?

Can anyone guide me regarding occlusion handling for tracking as well as in stereovision

[Dima Lisin]

Can you please be a bit more specific?

[Himanshu]

Yes,suppose i want to track a occluded car in a road scene how can i do that with 100% detection rate?

Answer 1

[Dima Lisin]

Hi Himanshu,

First of all, it is very rare in computer vision to see a 100% accuracy rate. Even the best detection or tracking algorithms still make some mistakes.

As far as your situation, more information would help. What algorithm do you use for object detection? What algorithm do you use for tracking? Is your camera stationary? Are there other objects in the scene besides cars?

For example, if your camera is stationary, you can use background subtraction to detect moving objects (e. g. vision.ForegroundDetector). Then you can use vision.KalmanFilter to track those objects using their motion. If the only moving objects in your scene are cars, then you should be able to track a car even if it is partially occluded, and even it is fully occluded for a short time.

[Himanshu]

I have attached one image i want to detect this car? Currently i am just working with images , i am not implementing it with camera for any real time system How can i detect car in this image as it is occluded by cycles I have both training and testing image dataset for this kind of problem I am using HOG for shape feature , i am getting the feature vector on matlab but due to lack of proper guidance i am unable to figure out what to do next after getting the feature vector?

[Dima Lisin]

Well, you should have a training set of images of cars and non-cars. Then you compute the feature vector for each image, and you train a classifier to distinguish between cars and non-cars. See this example. Then you write a detector, which slides a window across the image and uses the classifier to see if it contains a car.

The problem is that there will always be cases for which your classifier will be wrong. The case with a car occluded by bicycles is very difficult.

[Himanshu]

Ok so I have a training set of cars but in an image 80% is covered with car and rest 20% other things e.g building road etc. So firstly should i perform segmentation or is there any other way by which I can calculate the features of car and other things? So that i can differentiate between car and non car Secondly without making the car detection algorithm can I directly go for the problem of occlusion?

[Dima Lisin]

If the car takes up 80 of the image, that is actually a good situation. Typically, this is what a car-vs.-non-car classifier should be able to handle. However, occlusion is another matter.

Here's a thought. Bag-of-features approach is generally more robust to occlusion and clutter than a HOG-based classifier. That may be worth a try for you. See this example.

[Himanshu]

Ok i am working on this. Now I want to ask to handle object-object or car-car occlusion in a road scene , this is the first step i should start with which is image classification?

[Himanshu]

Secondly i am using this kind of images in which we have car only e.g simple cars of different companies for training so object classification is necessary