%0 Journal Article %A Saptharishi, Mahesh %A Oliver, C. Spence %A Diehl, Christopher P. %A Bhat, Kiran S. %A Dolan, John %A Trebi-Ollennu, Ashitey %A Khosla, Pradeep %D 2002 %T Distributed Surveillance and Reconnaissance Using Multiple Autonomous ATVs: CyberScout %U https://kilthub.cmu.edu/articles/journal_contribution/Distributed_Surveillance_and_Reconnaissance_Using_Multiple_Autonomous_ATVs_CyberScout/6621992 %R 10.1184/R1/6621992.v1 %2 https://kilthub.cmu.edu/ndownloader/files/12118481 %K Multiple autonomous vehicles %K dynamic path planning %K visual surveillance %K reconnaissance %K motion detection with image mosaics %K object classification %K moving object correspondence %X

The objective of the CyberScout project is to develop an autonomous surveillance and reconnaissance system using a network of all-terrain vehicles. In this paper, we focus on two facets of this system: 1) vision for surveillance and 2) autonomous navigation and dynamic path planning. 

In the area of vision-based surveillance, we have developed robust, efficient algorithms to detect, classify, and track moving objects of interest (person, people, or vehicle) with a static camera. Adaptation through feedback from the classifier and tracker allow the detector to use grayscale imagery, but perform as well as prior color-based detectors. We have extended the detector using scene mosaicing to detect and index moving objects when the camera is panning or tilting. The classification algorithm performs well (less than 8% error rate for all classes) with coarse inputs (20x20-pixel binary image chips), has unparalleled rejection capabilities (rejects 72% of spurious detections), and can flag novel moving objects. The tracking algorithm achieves highly accurate (96%) frame-to-frame correspondence for multiple moving objects in cluttered scenes by determining the discriminant relevance of object features. 

We have also developed a novel mission coordination architecture, CPAD (Checkpoint/Priority/Action Database), which performs path planning via checkpoint and dynamic priority assignment, using statistical estimates of the environment’s motion structure. The motion structure is used to make both preplanning and reactive behaviors more efficient by applying global context. This approach is more computationally efficient than centralized approaches and exploits robot cooperation in dynamic environments better than decoupled approaches.

%I Carnegie Mellon University