Automatic behavior recognition has been the topic of joint research by Noldus Information Technology and the Rudolf Magnus Institute for several years. The spin-off of this work is the EthoVision video tracking, motion analysis and behavior recognition system. Until now, this system is mainly used as an instrument to measure the behavior of animals on the basis of movement (using path-related parameters). To increase the capabilities of EthoVision to measure subtle behaviors of rodents, we are currently focussing our research on the automatic recognition of behaviors which are not related to movement. This approach is based on shape analysis. Examples of standard image features that describe the shape of an object are surface area, circularity, perimeter, moments, etc. In the cases of rodents, however, these standard object shape features do not provide enough information for automatic behavior recognition. Therefore we have developed new algorithms for the extraction of model-based features. Since our first report on this work [2] we have improved the detection of these features and have extended the range to include the following: X,Y coordinate of the head point (snout), tail base (point where the tail is attached to the abdomen), corrected center of gravity and several additional features derived from these three points. These new features together with the features describing the shape of the object are being used for automated behavior recognition.

Automatic scoring of behaviors can be seen as a form of pattern classification, the process which assigns classes to signals that are derived from objects using a sensory system. In pattern classification there are, in general, two approaches used for classification. One is the use of a statistical classifier, the other is the use of neural networks. At Noldus Information Technology we are investigating the statistical approach, while our project partners at the Rudolf Magnus Institute study the use of neural networks [1]. For statistical pattern classification we use a Bayes classifier. This classification method is based on the following two prerequisites: (1) the damage involved when an object is classified incorrectly can be quantified as cost, and (2) the expectation of the cost (known as risk) is acceptable as optimization criterion. Before actual classification of the data, the classifier has to be trained. During training, information about the distribution of the features for the different classes is calculated. The data used for training consists of the features describing the object plus the behaviors as scored by a human observer. Once the classifier is trained it is ready to use for classifying new data.

It is expected that especially standard behavioral tests such as the elevated plus maze a nd open field experiments can be further automated with this kind of technology. Examples of some typical behaviors that may be recognized automatically are head dip, head raise, stretched attend, rearing, grooming and crouching.

The research is still in progress; the results obtained so far will be presented at the conference. This work is carried out with financial support from Eureka project no. EU88011.

References

1. Rousseau, J.B.I.; Melder, W.; Costa-Florencio, C; van Lochem, P.B.A.; Gispen, W.H; Spruijt, B.M. (1998). Classification of rat behavior by a neural network. This volume.
2. Smit, J.; Rousseau, J.B.I.; van Lochem, P.B.A.; Plakke, R (1996). Automatic recognition of behavioral patterns in rodents using digital imaging. Abstr. Measuring Behavior '96 (Utrecht, 16-18 October 1996), 95-96.