Research into Computer Vision
Lately I have been doing some thinking about research into computer vision, specifically in the area of motion detection and object recognition. I want to get a handle on the basics before I make a decision into whether or not I would like to study this as part of a Masters of IT. As part of this I have come up with an idea to help me along the way. My approach is going to be two fold:
1) Develop a system to simplify objects in a video stream.
2) Develop a system to classify and/or recognise objects of interest based on their attributes and persist these objects across frames.
1. is obviously dependent on 2. so I am starting with 1.
I currently have under development an application that takes a video stream from a webcam/IP Camera and simplifies the image. Basically I am working on the idea that objects boundaries are found be determining areas of rapid pallette change in a given video frame.
The video simplify system basically works this way.
1. Get a frame from a video stream
2. Pre-Process the frame with some despeckling routines to remove the noise.
3. Pass the frame through two filters, both do the same thing put with different parameters. The filters to the following:
Edge detect the image
Binarise the edges based on a threshold
Skeletalise the binarised edges
The process looks like this:
So from a video stream we now have a simplified wire frame model of the view. Note this is in real-time ( about 11 fps on my mobile centrino ).
I am getting quite a bit of noise in the images because I am using a really cheap web camera for this. I will definitely need to upgrade if I want to take this further.
The other part of this is object persisitance, the reason I want this is because if you can persist (recognize and remember) an object across more that one frame then it is possible to build a system that can judge distances with a single camera. If you can judge distances then you can navigate obstacles. Will post more on this when I have done some more.
1) Develop a system to simplify objects in a video stream.
2) Develop a system to classify and/or recognise objects of interest based on their attributes and persist these objects across frames.
1. is obviously dependent on 2. so I am starting with 1.
I currently have under development an application that takes a video stream from a webcam/IP Camera and simplifies the image. Basically I am working on the idea that objects boundaries are found be determining areas of rapid pallette change in a given video frame.
The video simplify system basically works this way.
1. Get a frame from a video stream
2. Pre-Process the frame with some despeckling routines to remove the noise.
3. Pass the frame through two filters, both do the same thing put with different parameters. The filters to the following:
Edge detect the image
Binarise the edges based on a threshold
Skeletalise the binarised edges
The process looks like this:
Original Image
Edge Detected
Thresholded and sketalised
Merged results
So from a video stream we now have a simplified wire frame model of the view. Note this is in real-time ( about 11 fps on my mobile centrino ).
I am getting quite a bit of noise in the images because I am using a really cheap web camera for this. I will definitely need to upgrade if I want to take this further.
The other part of this is object persisitance, the reason I want this is because if you can persist (recognize and remember) an object across more that one frame then it is possible to build a system that can judge distances with a single camera. If you can judge distances then you can navigate obstacles. Will post more on this when I have done some more.
posted by TheHojuSaram at
4:13 pm
0 Comments:
Post a Comment
<< Home