Monitoring behavior of individuals in crowded scenes

A.J. Bulpitt1, R.D. Boyle1 and J.M. Forbes2

1School of Computer Studies, University of Leeds, Leeds, United Kingdom
2Centre for Animal Sciences, Leeds Institute of Biotechnology and Agriculture, University of Leeds, Leeds, United Kingdom

Much research in the area of automated monitoring is focused on tracking a small number of individuals, often in well-controlled environments. We present a method for tracking large numbers of objects simultaneously in open and crowded scenes over extended periods. Our application is monitoring the behavior of broiler chickens for assessing health and welfare. The application offers several challenges:

In order solve these problems our approach employs methods for modeling both the shape and gray level variations of the objects in the scene combined in a hybrid Active Shape Model (ASM)/Condition Density Propagation (Condensation) framework [1,2].

Modeling the expected shape of individuals in an ASM, permits the identification of individuals in a crowd rather than only identifying groups. The Condensation framework provides a robust and efficient method of tracking the individuals over a cluttered and evolving background. This approach also overcomes the problem of the variable short term behavior of the birds, as it removes the need for a good a priori model of behavior required by a Kalman filter based solution. Other methods for identifying individuals such as that of Sergeant [4] do not attempt to model the shape or motion of the objects but solve a correspondence problem between frames. This becomes difficult to solve with the condition of an open scene, as the number of birds may not be consistent between frames.

The problems of poor contrast between the birds and the background and poor discriminating boundaries are resolved by using a measure of fuzz affinity [3] to test the homogeneity of regions in the image. The robustness of the tracker is further improved by modeling the gray level intensities of individuals in the scene, which is automatically updated to adapt to changes in lighting intensity.

The system has currently been evaluated over a 5-minute period during which 93 individuals entered the scene and up to 30 birds were required to be tracked simultaneously. Current results demonstrate over 90% of the birds can be tracked successfully over this period (Figure 1). The performance achieved is certainly high enough for analyzing the behavior of the birds and overcomes many of the problems associated with human observation over extended periods including intra- and inter-observer errors.

Figure 1. Result for five video frames and trajectories produced

References

  1. Cootes, T.; Taylor, C.; Cooper, D.; Graham, J. (1995). Active Shape Models - their training and application. Computer Vision and Image Understanding, 61 (1), 38-59.
  2. Isard, M.; Blake, A. (1996). Contour tracking by stochastic propagation of conditional density. Proc. 4th European Conference on Computer Vision, 343-356.
  3. Jones, T.N.; Metaxas, D.N. (1998). Image segmentation based on the integration of pixel affinity and deformable models. Proc. Conference on Computer Vision and Pattern Recognition, 300-337.
  4. Sergeant, D.M.; Boyle, R.D.; Forbes, J.M. (1998). Computer visual tracking of poultry. Computers and Electronics in Agriculture, 21 (1), 1-18.

Paper presented at Measuring Behavior 2000, 3rd International Conference on Methods and Techniques in Behavioral Research, 15-18 August 2000, Nijmegen, The Netherlands

© 2000 Noldus Information Technology b.v.