Moving object classification

edit

The aspect ratio of cars and moterbikes may be relatively less. But, cars will have more width (and length) You can check the length and width of each blob and set a threshold to classify the vehicles.

Also, you can check the blob properties in an ROI (the junction) to make it more effective. The blob area and aspect ratio wouldn't vary that much within the ROI

Have a look at the cascade classifier to effectively "recognize" cars and motorbikes. This classifier is able to perform multi-scale recognition, so the difference in appearance shouldn't be a problem. OpenCV comes with an already trained car detector, but I don't know if it is suitable for your needs.

In a general approach, you don't necessarily perform object tracking before the recognition, but as you already have it and it could speed up a lot! your recognition stage (only try to recognize the detected object), you should try to keep it.

If you don't know how to start with the cascade classifier, have a look at these tutorials or the chapter 5 of the book named "OpenCV 3 Blueprints". They will give you a lot of indication on how to start, and (in the book) how important it is to collect a good dataset to fulfill your needs.