Dmitriy Korchev

Detecting small, low-contrast moving targets in infrared video produced by inconsistent sensor with bad pixels

Abstract. This paper addresses the problem of finding small and low-contrast moving targets in infrared (IR) video sequences produced by sensors with inconsistent parameters, such as intensity offset and gain as well as bad pixels. This sensor variability makes it difficult to apply methods based on frame registration using simple pixel differences. Our proposed algorithm uses regression to normalize the variations of intensity offset and gain between compared registered frames. A statistical criterion is used to calculate the threshold for the difference between normalized intensities of two frames. The algorithm for finding the differences between frames is also used to create a bad pixel mask either on- or offline. This mask is essential for the reduction of false detection rates. Our experiments show that this approach produces good results and can be used for detection of small, low-contrast targets in high dynamic range IR data. The proposed algorithm also produces good results for detecting moving targets in cases when objects are occluded by sparse vegetation.

Multisensor ISR in geo-registered contextual visual dataspace (CVD)

Current ISR (Intelligence, Surveillance, and Reconnaissance) systems require an analyst to observe each video stream, which will result in analyst overload as systems such as ARGUS or Gorgon Stare come into use with many video streams generated by those sensor platforms. Full exploitation of these new sensors is not possible using todays one video stream per analyst paradigm. The Contextual Visual Dataspace (CVD) is a compact representation of real-time updating of dynamic objects from multiple video streams in a global (geo-registered/annotated) view that combines automated 3D modeling and semantic labeling of a scene. CVD provides a single integrated view of multiple automatically-selected video windows with 3D context. For a proof of concept, a CVD demonstration system performing detection, localization, and tracking of dynamic objects (e.g., vehicles and pedestrians) in multiple infrastructure camera views was developed using a combination of known computer vision methods, including foreground detection by background subtraction, ground-plane homography mapping, and appearance model-based tracking. Automated labeling of fixed and moving objects enables intelligent context-aware tracking and behavior analysis and will greatly improve ISR capabilities.

Algorithm for detecting important changes in lidar point clouds

Protection of installations in hostile environments is a very critical part of military and civilian operations that requires a significant amount of security personnel to be deployed around the clock. Any electronic change detection system for detection of threats must have high probability of detection and low false alarm rates to be useful in the presence of natural motion of trees and vegetation due to wind. We propose a 3D change detection system based on a LIDAR sensor that can reliably and robustly detect threats and intrusions in different environments including surrounding trees, vegetation, and other natural landscape features. Our LIDAR processing algorithm finds human activity and human-caused changes not only in open spaces but also in heavy vegetated areas hidden from direct observation by 2D imaging sensors. The algorithm processes a sequence of point clouds called frames. Every 3D frame is mapped into a 2D horizontal rectangular grid. Each cell of this grid is processed to calculate the distribution of the points mapped into it. The spatial differences are detected by analyzing the differences in distributions of the corresponding cells that belong to different frames. Several heuristic filters are considered to reduce false detections caused by natural changes in the environment.