Tracking Objects Across Spatially Seperated Cameras with Non-overlapping Views

This research project looks at an algorithm by A. Gilbert and R. Bowden that tracks objects across cameras with non-overlapping views using three cues: color, object size, and movement. We focus only on the performance of the color cue. A frame of each view of a camera is an image that is made up of small blocks termed pixels. Each pixel has a color that is represented by red, green, and blue values that range from 0 to 255. Our setup uses three spatially separated cameras. Objects are tracked manually in these cameras, and red, green, and blue histograms, a color descriptor, are produced for each. A color histogram is created by dividing each color into a number of bins; we use 6 bins that evenly divide the range 0 to 255. The red, green and blue value of a pixel is distributed into its correct histogram and bin. These histograms are compared to each other within camera and across cameras using histogram intersection. Because of different lighting and camera settings, the color description can significantly change for an object in different camera views. Therefore, we evaluate the use of a transform matrix to remove the color difference across the cameras, a process known as color calibration. Comparing the calibrated to uncalibrated color histograms across cameras, we see a significant gain. In conclusion, we see that after the color calibration, the color description of an object across cameras is very similar to an object being compared within camera.