Military Detection & Tracking

militarytracking pic.jpg

Project Info

Title Projection Model Snakes for Tracking Using a Monte Carlo Approach, Clutter-Reducing, Scalable Image Processing Methods for Target Tracking
Goal Develop and implement robust, effective tracking algorithms for military target tracking
Funding United States Army Research Office


To effectively track military targets while preventing aimpoint drift in the presence of significant clutter and scale changes. Development of a robust tracking system will improve missile accuracy to reduce cost and collateral damage.


Efforts to improve military target tracking are necessary because of the complex requirements of such a tracking system. Missiles must track and destroy moving targets obscured by clutter and image noise. In addition, due to the descent of a missile upon its target, a missile tracking system must be capable of handling significant scale changes and changes with orientation to the object in three dimensions. We demonstrate the usefulness of established methods and some novel approaches to solving this difficult tracking problem.

The tracking data, obtained from the Army Research Office and the Army Aviation and Missile Command (AMCOM) at Redstone Arsenal, consist of simulated air-to-ground sequences using forward-looking infrared camera technology. These sequences demonstrate the clutter, noise, scale, and three-dimensional registration issues that complicate the tracking process.

Because of the clutter, noise, and scale, traditional correlation and centroid tracking algorithms have proven effective but not truly robust. In addition, these trackers are susceptible to aimpoint drift, meaning that the missile may hit the target, but not in the right place. We employ a novel method called projection model snakes, which are good for tracking rigid bodies such as tanks or armored personnel vehicles. These projection model snakes track target boundaries by finding a geometric transformation which describes the movement of a rigid target boundary from frame to frame. We consider both affine and projective models for the projection model snakes to account for the range of scales encountered during closure sequences, where the viewer progresses from being far away from the target to being very close. The projective model is appropriate when the viewer is very close to the target and parallel lines may appear to skew and meet at some point in the distance.

The affine snake is a variation of this concept where we use an energy functional to find the parameters of the affine transformation that maximizes the gradient that the contour boundary lies on. We also incorporate a smoothness constraint into the energy functional that constrains the affine transformation to be smooth and gradual, preventing drastic evolutions during tracking. The affine parameters that achieve the minimal energy value determine the affine transformation (rotation, scaling, and translation) for the location of the target boundary in the current frame. The tracker uses a generalized gradient vector flow (GGVF) method to find the image gradient force. Implementation of the projective version of this algorithm is similar. Results are available below in avi format showing tracking of a B2 bomber sequence with the non-rigid snake and affine snake.

  • Non-rigid snake [.avi]
  • Affine snake without smoothness [.avi]
  • Affine snake with smoothness [.avi]

We also extend the use of projection models by combining them with sequential Monte Carlo methods for stochastic tracking. The basic idea of the affine Monte Carlo algorithm is to apply random perturbations to the translation, rotation, and scale of many contours and weight each by its likelihood of actually capturing the target. This likelihood is based on the mean gradient around the curve, so if the contours lie on high gradient (target boundaries) they have high weights. The estimated target boundary contour is then found as a combination of the likelihood-weighted random paths. This algorithm is extremely effective in clutter, noise, occlusion and in tracking dynamic targets. An example below demonstrates the effectiveness of the affine Monte Carlo tracker over non-rigid and affine snake energy minimization algorithms in tracking a synthetic military sequence. The affine Monte Carlo method tracks successfully through occlusion while the non-rigid snake algorithm collapses on itself and the affine snake actually switches targets. This demonstrates one of the benefits of a stochastic tracker because it is able to escape local minima, unlike most greedy energy minimization algorithms.

  • Non-rigid snake tracking a humvee [.avi]
  • Affine snake tracking a humvee [.avi]
  • Affine Monte Carlo method tracking a humvee [.avi]

Nonlinear and morphological filtering methods have also proven useful for improving tracking results. In particular, area morphology, morphological anisotropic diffusion, and morphological locally-monotonic filters improve the efficacy of the affine snake and other trackers. Below is an example of a GGVF with and without the use of morphological anisotropic diffusion. The reduction in the number of spurious edges from clutter and noise is highly evident.

militarytracking fig1.jpg
Fig 1. GGVF.

militarytracking fig2.jpg
Fig 2. GGVF with morphological anisotropic diffusion.

Examples of tracking with and without the use of area morphology are shown below.

  • Jeep with no filtering [.avi]
  • Jeep with area open-close [.avi]

Mukherjee's Revenge is a graphical user interface which we developed for testing and visualization of the affine snake and the various filtering techniques. Below is a screenshot of Mukherjee's Revenge acquiring and preparing to track a military target.

militarytracking fig3.jpg
Fig 3. Mukherjee's Revenge screenshot.

Future research lies in three-dimensional registration of two images from different perspectives, predicting camera location given the range and altitude information to the target, and automated initialization of the snake polygon.


Topic revision: r3 - 30 May 2014, AndreaVaccari
©2018 University of Virginia. Privacy Statement
Virginia Image and Video Analysis - School of Engineering and Applied Science - University of Virginia
P.O. Box 400743 - Charlottesville, VA - 22904 - E-Mail