MULTIPLE OBJECTS TRACKING USING BLACKBOARD, PYRAMID AND TRANSPUTERS

Abstract

This thesis describes a computer vision system that is able to track up to six objects simultaneously. The system implemented is based on the model proposed by TAN and MARTIN on multi-resolution imagery and distributed processing. The system adopts a pyramid image structure. It scans for objects using the lowest resolution image al the top of the pyramid, and once an object is found, it descends towards the base of the pyramid, getting progressively higher resolution images as it moves down. Because the shape of objects may look alike al low resolution, the top level processing only serves to indicate that there may be an object of interest. The identity of the object is confirmed only at the base of the pyramid - the highest resolution. Once the object's identity is confirmed, the system will track the object at the lowest resolution level again. The system has six "agents" responsible for detecting, identifying and tracking objects. These six agents obtain image segments from a video processor which is capable of providing images of different resolution according to the levels of the pyramid at which various agents arc working on. The six agents operate asynchronously and independently. To effectively manage the six agents, the concept of a blackboard is also implemented in this system. The blackboard serves as a global database that records attributes of objects reported by agents. The blackboard being visible to all agents also prevents an agent from pursuing an object which has already been reported by an earlier agent. The information in the blackboard also provide an instantaneous description of the entire dynamic scene. A 10 transputer network has been chosen as the platform for implementing the tracker. One transputer serves as the blackboard, one acts as the video processor and the other eight work as six agents and two communicators. The communicators are "postman" introduced to regulate message traffic. This is because each transputer has only four external links, it is therefore impossible for the blackboard to "talk" to the six agents without the help of the communicators. The use of transputers reveals a number of issues in parallel processing. Among these are bottle necks at common access areas and the need for semaphore to protect common data. The transputer platform also provides an opportunity to collect empirical data on actual performance. From the data collected, the overhead spent in communicating with blackboard was found to be negligible. This object tracking system is proposed for use in a number of surveillance applications ranging from Unmanned Air Vehicles, Harbour surveillance to traffic monitoring. There is also an ongoing project using the same concept to achieve a better understanding of raw data provided by multiple sensors.