tldr: For 2D image, you need the "reference object" and the "perception of depth and distance"

To tackle these kind of problems, the reference object is needed. The reference object is used for calibrating the object in the image dimension to the real world dimension. The reference object can be an item that has specific lenght, width and/or height. For an example, the lane marking on the road can be used as the reference point to classified the vehicle type( however the vehicle type itself is quite unique but if you have trouble on computational power, classification from comparing the referece object might be faster and template matching or other fancy CNN classification). Plus, the lane marking is stationary painted with the same gap. therefore the perception of depth and distance is also given.

You can try search for "2d object classification" and "machine learning". currently, i am very satisfy with tensorflow for my classification problem. But it requires tremendeous workload on data labeling.

Good luck and have fun,

George

https://medium.com/@tifa2up/image-classification-using-deep-neural-networks-a-beginner-friendly-approach-using-tensorflow-94b0a090ccd4

hi,

yes it is a chalenging task as you need some kind of semantic image processing, you may think of creating a database which includes the objects that you are going to identify in the scene and their sizees.

Using depth imaging is another option in which you can identify the location of each objects

best of luck

You can use moments as preprocessing method. Then, NN or a another approach for classification