Crazor's profile - activity

I need to recognize an object that is larger than the camera FOV. It is moving through the frame from side to side at close proximity to the camera. I want to trigger an event when the object is centered in front of the camera, so I need an approximate center point. Right now, I'm using an lbp cascade that works very well when the whole object is in the frame. My guess would be to train several cascades that recognise specific parts of the object, but I'm interested in other ideas.

Something is wrong either with your training data (samples/negatives) or with your training procedure. HR 1/FA 0 in the first stage means that the classifier locates the object with 100% accuracy after running only a single stage, with zero false alarms on negative images. With correct training, you will get HR 1/FA 1, with FA slowly decreasing with number of selected features. If FA drops below the margin you specified (default 0.5), the next stage is trained. This takes anywhere between minutes or hours to days, depending on the kind of classifier (Haar/LBP), number of sample images/negatives and number of stages to train.

This is a good tutorial to get started: http://coding-robin.de/2013/07/22/tra...

Sorry, I didn't get notified of the comment.

I can't post images. But even without one, this question should be answerable. I'm getting a vector of lines that follow the object's contours, and I'd like to run an algorithm that learns to classify if the lines are from the object I'm looking for or not.

I'm building a system that analyzes a video feed. I need to decide wheter a specific object is visible and where it is. Because of the structure of the object and the variable circumstances of the scene, I came up with the following: MOG2 for BG subtraction, Open/Close to get rid of noise, Canny, then probabilistic Hough. I've tuned all the parameters to get me a reliable distinction between the object I'm looking for and all the other random stuff around. Basically, the object I'm looking for lights up with a nice bunch of line segments, while I get the ocassional one or two lines from the surroundings.

How would I now go about deciding if and where the object is, based on the list of line segments? Note that the object moves through the scene, and I need to recognize the object and it's approximate position, so that I can later trigger an output depending on where my object is.

I'm thinking about ML/NNs/SVMs, but I'm not familiar with any of those techniques. I would be happy if someone could point out what would be the best way to continue here, so I can go read up on what to implement next.