Predictive Tracking Using Kalman Filtering

Author: Ben McEwen

Hi, I’m Ben. I am a student in my final year of Mechatronics Engineering at the University of Canterbury. This semester I have been working with the Cacophony Project to explore improved ways of tracking animals in video footage for better predator recognition and for the elimination of invasive predator species.

Improved Tracking

Through the use of a Kalman filter, there has been a significant reduction in noise and better handling of occlusion when following an animal's movement. Kalman filters are a powerful tool which can be used to predict the current and future state of a system using the weighting between the sensor input uncertainty and previous estimate uncertainty. Kalman filters are computationally efficient and well suited to embedded system applications.

Occlusion commonly occurs when animals move behind objects resulting in the animal’s form being broken up. In the specific case of possums being observed via thermal video, a common issue is that the body and tail of the animal are displayed as two separate heat signatures. This introduces noise when tracking the centroid of the animal and causes sudden jumps in the coordinate of the centroid, shown by the blue line in the following figure. The red line shows the much smoother estimated position produced by the Kalman filter.

Predictive Tracking

The Kalman filter can also be used to predict the future position of an animal. The velocity of the animal was predicted and then filtered using a moving average filter. The time step can be increased to predict further in front of the animal up to around 3-5 seconds depending on the size and speed of the animal. This could potentially be used to assist with targeting pest species for toxin delivery. The idea of using an automated turret capable of firing projectiles filled with small doses of toxin has been discussed in previous blog posts.

This software provides a method for accurately predicting the future position of animals moving at varying speeds and with different movement patterns. The following video shows a thermal recording of a possum. On the left is the processed recording showing the estimated centroid of the animal, the movement vector and the predicted position. The original recording is shown on the right, with the predicted position overlaid.

Summary

Initial results look very promising. There has been a reduction in noise and better handling of occlusion when tracking animal movement and the future position of animals can be accurately predicted. In the future, potential improvements could be made to the prediction range using non-linear Kalman filters such as the Extended Kalman filter (EKF) and Unscented Kalman filter (UKF).

For further details about Ben's work, see his paper:Predictive Animal Tracking for Invasive Species Identification and Elimination.