In version 4 the classification model has been updated with new categories and given a general overhaul. Previously the model only differenced between cars and people. Now it will categorise cars, people, vans, bikes, motorbikes, trucks, buses and miscellaneous for anything that doesn’t match any of the previously mentioned categorise.
Importing a video using the Event Detection and Classification is the same as before, just select that algorithm from the list in the analysis algorithm page of the import wizard. There are advanced setting for this algorithm including the ability to select the model used manually and change the inflate sprite percentage. By default the latest model is selected.
Inflate Sprite Percentage
The input into to the object classifier is a region, or sub-image, extracted from a video frame where motion has been detected by the Event Detection algorithm. Ideally, the extracted region is perfectly localised around the moving object, however, in many cases, the detected motion event will be in-and-around the moving object. With this in mind, we have introduced an Inflate Sprite Percentage option which expands the extracted region so that it is more likely to capture the source of the motion, thus leading to better classification results.
The suggested range of values for the Inflate Sprite Percentage option is between 0% (for videos with accurately localised moving objects) and 50% (for videos with poorly localised moving objects).
Once you’ve imported a video with the classification algorithm you’ll be presented with the Select object types page in the filter wizard. Here you can create a filter for just a specific type of object and only events containing those objects will appear on the timeline. If the video hasn’t been processed with the classification algorithm then this page will not appear in the wizard.
Classification objects still appear in the Grid tab as they’ve always done but now with two new options.
Every object has a confidence level attached to it. This level is how confident the model is that it’s correctly categorised the object. The slider in the filter section allows you to quickly hide any objects from the grid that are under a certain level of confidence.
You can also sort by confidence too. By default this is done in descending order so the objects the model is most confident about appear on the top.
For more information about getting the most out of the classification algorithm please see the Kinesense Deployment Consideration document that appears in the Help tab in the software.