|PROJECTS AT VISL FINISHED IN 2005|
The final goal of this project was to develop an autonomous vehicle that uses a small camera and a Pocket PC computer to track and follow a walking person. In our work we dealt with finding a suitable algorithm for this mission. Several possible algorithms were examined theoretically, three of them were implemented in MATLAB and practically checked on sample movies.
The objectives of the algorithm:
Few basic assumptions used in our methods:
The three algorithms that we implemented and tested are:
1. Movement estimation base block comparison
In this method each frame is divided to blocks of 16X16. For each block we then try to estimate movement by correlation with the near area in the following frame. With these results we produce a movement vectors diagram, which indicates which blocks have moved, and where to, so we can determine to which direction the figure moved.
Movement estimation between 2 frames
2. Feet tracking based on recognition of diagonals in the image
This algorithm was motivated from an article explaining a method of recognizing people walking in an urban environment. The method is based the assumption that in an urban surrounding most lines should be straight lines. This assumption may enable us to filter these lines and remain with the legs of walking persons. We implemented such an algorithm and tried to track pedestrian legs by this method.
3. Tracking according to colors characteristics
In this method, we first recognize the motion in the movie's frames by differentiating two concurrent frames. A new legal target is defined by as the most significant block of "moving pixels" in the current frame. In each frame , the algorithm draws a window around the " target". The window is adaptive so that its size changes when the "target" gets closer or farer to the camera. Once we found a target we analyze the color histogram of this block and determine the most dominant colors in this area. In the next frames we search for the areas that contain the same colors. When a target is lost, we search for new targets as described before.
In the above image, the red areas are the most common colors of the target , after filtering RGB components , we calculate the center of mass of these areas .
Sample movie of succesful tracking: [movie]
The algorithm was ineffective due to the following reasons:
This algorithm may work well under certain circumstances, but it has some disadvantages:
3.Tracking based on colors recognition
This algorithm worked the best for our example movies: