Tracking Dismounts With Feature Matrix Based Multi-scale Intrinsic Motion Segmentation Framework (FM-MIMS)

Abstract

Tracking dismounts using wide-area motion imagery (WAMI) from the air is a challenging problem. Such an algorithm has many applications on the battlefield, search and rescue, law enforcement, and more. This task is however nontrivial. Small target sizes and difficult to identify features make it extremely challenging to reliably detect people on the ground. Based on the Multi-scale Intrinsic Motion Structure framework (MIMS), this algorithm proposes an extension to the MIMS framework to allow simple feature identifiers to help improve the rate of successfully identifying a dismount. The Feature Matrix Based Multi-scale Intrinsic Motion Segmentation (FM-MIMS) does so by encoding size and motion features from a small target detector, and an optical flow detector. The features are used to drive the predictions made by a tensor voting algorithm that is then segmented to identify the target. Given the computationally intensive nature of tensor voting, this thesis proposes incremental changes to it as well as a layer of image prepossessing to allow for a robust method to detect dismounts. Feature extraction would allow the MIMS framework to extend its 4-9pixel target size to better detect larger targets in the 15-20pixel size range. At this scale dismounts are still very limited in identifiable features however, appendages such as arms and legs are visible. FM-MIMS intends to extract as much data as possible to increase the reliability of the detector.