Zoom techniques for achieving scale invariant object tracking in real-time active vision systems

In a surveillance system, a camera operator follows an object of interest by moving the camera, then gains additional information about the object by zooming. As the active vision field advances, the ability to automate such a system is nearing fruition. One hurdle limiting the use of object recognition algorithms in real-time systems is the quality of captured imagery; recognition algorithms often have strict scale and position requirements where if those parameters are not met, the performance rapidly degrades to failure. The ability of an automatic fixation system to capture quality video of an accelerating target is directly related to the response time of the mechanical pan, tilt, and zoom platform—however the price of such a platform rises with its performance. The goal of this work is to create a system that provides scale-invariant tracking using inexpensive off-the-shelf components. Since optical zoom acts as a measurement gain, amplifying both resolution and tracking error, a new second camera with fixed focal length assists the zooming camera if it loses fixation—effectively clipping error. Furthermore, digital zoom adjusts the captured image to ensure position and scale invariance for the higher-level application. The implemented system uses two Sony EVI-D100 cameras on a 2.8GHz Dual Pentium Xeon PC. This work presents experiments to exhibit the effectiveness of the system