Whats the best way to do multiple object recognition/classification in real time using OpenCV and GPU (CUDA)?

edit

I’m figuring out a ways to recognition multiple object in real time like the image shown here. To speed up the process i thought of utilizing GPU.

Do i need to go for image segmentation then try to recognition or try to implement it fully in Convolutional Neural Networks (CNN). i’m willing to learn new things and try them out. How do I go about it? or how to start? Any step by step process or links is appreciated.