Onboard and parts-based object detection from aerial imagery

Abstract

The almost endless amount of full-motion video (FMV) data collected by Unmanned Aerial Vehicles (UAV) and similar sources presents mounting challenges to human analysts, particularly to their sustained attention to detail despite the monotony of continuous review. This digital deluge of raw imagery also places unsustainable loads on the limited resource of network bandwidth. Automated analysis onboard the UAV allows transmitting only pertinent portions of the imagery, reducing bandwidth usage and mitigating operator fatigue. Further, target detection and tracking information that is immediately available to the UAV facilitates more autonomous operations, with reduced communication needs to the ground station. Experimental results proved the utility of our onboard detection system a) through bandwidth reduction by two orders of magnitude and b) through reduced operator workload. Additionally, a novel parts-based detection method was developed. A whole-object detector is not well suited for deformable and articulated objects, and susceptible to failure due to partial occlusions. Parts detection with a subsequent structural model overcomes these difficulties, is potentially more computationally efficient (smaller resource footprint and able to be decomposed into a hierarchy), and permits reuse for multiple object types. Our parts-based vehicle detector achieved detection accuracy comparable to whole-object detection, yet exhibiting said advantages.