Chapter 5 Traffic Anomaly Detection

I can split 3 main tasks: - Detection - Merge them, remove noise bounding box (bbox) - Postprocessing like tracking back time, delete redundant object, …

5.1 Detection

This is a real problem, because how to pick what kind of models? A big one to detect all kind, split them to detect small car, what about weather, or nightime…

When working on this I try to use one big models, train in many images with small car, large car, night, day, … but in the end it happened to do a poor job. I have to lower the theshold inorder to take a risk of false positive, just to make it find a little tiny one in the corner. I was greedy, the cost is high amount of recall, to many false positive that I couldn’t fix in postprocessing stage. I really need to split the region to apply which config instead of use a whole one and hope for it to adapt many situations.