Knowledge Base

Behavior Recognition: Requirements

Species

• The Behavior Recognition function has been developed and tested for tracking rats and mice, not other animal species.

Number of Arenas

• When tracking live: one arena.
• When tracking from pre-recorded video files: one to four arenas.

Subjects

• If you use multiple arenas, the subjects must have similar size. The apparent size of the subjects is specified in the Detection Settings. That size is used for all arenas.
• The subject's length (nose to tail base) in the video image must be at least 60 pixels for rats, and 55 pixels for mice. In all cases its length must not exceed half the arena size.
• The subject's fur must be of uniform color. Hooded animals like Lister and Long-Evans rats have not been tested with Behavior Recognition.
• For rats: the subject must be older than approximately three weeks. In all cases the subject must be able to walk.
• In the Detection Settings, make sure that the subject's tail is not detected. To do so, use the erosion filter. See Advanced Detection Settings: Subject Contour.
• Behavior detection is much depending on the subject age and size. For Wistar rats of age 3–5 weeks, the behavior size setting can be used about one week; from age 5 weeks on, it may be used two weeks. We advise to create separate Detection Settings for different age classes.

Camera and Video

• The camera is placed above the subject, and provides a top view of the test setup.
• Because video compression introduces artifacts, live video is preferred over pre-recorded video.
• Video frame size (video resolution) must be greater than 352 x 288 per arena.
• The frame rate of the recorded videos can only be a multiple of the sample rate that you intend to use. For example, for a sample rate of 25 samples per second, the camera/video file frame rate should be 25 or 50 etc. For a sample rate of 30, set the frame rate to 30 or 60.
• The video image must not be overexposed. Details of the subject's fur must be visible.
• If you use pre-recorded video, compression should be as lossless as possible. Reducing compression by decreasing the GOP size (that is, the number of video frames between two full frames) produces better results than increasing video resolution. Values of GOP size of 10–15 worked well, however the lower the better. Reducing GOP size, however, increases file size.
• To check that your video file has artifacts, create a separate experiment with Activity analysis enabled. Set the Activity threshold to a minimum. See Activity Settings. If the video contains artifacts, you should see purple pixels appearing with a regular rhythm. See also Troubleshooting: Behavior Recognition.
• Try to avoid evident fish-eye effect (barrel distortion) in the video image, which hampers behavior recognition. If necessary, use a lens with lower distortion.

Test Environment

• The test setup must be simple, for example a home cage (without bars) or an open field with no objects around.
• The animal should not climb on top of objects like a shelter. When the animal climbs on top of an object, it gets closer to the camera, and its apparent size changes significantly. This may bias results. See also Behavior Recognition: Data, Performance and Accuracy.
• Lighting must be from the top, not backlighting. Slow changes in lighting in the course of the test are not a problem, but moving spotlights reduce reliability of detection.
• Avoid reflections on the walls.

Tracking Settings

• In the Experiment Settings, you must select Contour-based as Body Point Detection Technique.
• In the Detection Settings, the sample rate must be set between 25 and 31 samples per second.

Further Recommendations

• During acquisition, keep DDS (Detection Determines Speed) selected.
• After acquiring data, visualize the detected behaviors in the Integrated Visualization and check whether no gaps between scored behaviors occur. If gaps occur, which are not due to the fact that the subject is not found, they could be the result of video frames not being analyzed due to high processor load. To decrease processor load, lower the video resolution. Also, try not to use Differencing as detection method, and any of the methods named For Occlusions as the tracking method. See Advanced Detection Settings: Method.

Behavior Recognition: Data, Performance and Accuracy

Data

With the Rat/Mouse Behavior Recognition functions, the following behaviors can be automatically detected: digging (only for mice), drinking, eating, grooming, hopping (for mice), jumping (for rats), rearing (unsupported and supported), resting, sniffing, twitching (for rats), and walking. For the definition of those behaviors, see Behavior Recognition.

For each video frame, behavior is detected by analyzing a number of frames before and after the current one.

• Behavior data are available from the start of the track, to about 13 seconds before the end of the track when using default Trial Control Settings, which has an additional condition for starting the track.
• When the Trial Control Settings does not have the special condition mentioned above, behavior is not scored in the first 3 seconds of the track.
• When samples are missed or the subject is not found for a maximum of 0.4 seconds, the subject's position and size is interpolated using the adjacent frames.
• When samples are missed or the subject is not found for more than 0.4 seconds, or when the subject enters a hidden zone, the behavior detector is reset. This means that behavior is not scored some time before the first missing sample (or before entering the hidden zone) and after the last missing sample (or after exiting the hidden zone). The behavior is set to Unknown.

We advise you to adjust the duration of the trial in the Trial Control Settings. For example, set 6 minutes of track duration to get at least 5 minutes of automatically detected behaviors.

The first part of the track is scored as Unknown only if you start tracking as soon as the animal is detected in the arena. If your Trial Control rule includes a condition that waits 3 seconds after the animal is detected, then you make sure that behavior is scored from the start of the track. We recommend using the default Trial Control settings.