The optical detection and tracking of vehicles is a long-researched area. The algorithms are continuously evolving, but still there is no standard method to solve this problem. The statistical analysis of vehicles plays a crucial role in transport infrastructure development. The well-founded developments around the world allow for savings of billions of euros annually. During my project I had to create a system which can solve this problem according to our needs.
This thesis mainly summarizes deep learning based technologies that we use at Idaso to find a solution to the problem. I present a series of motion detector based experiments where I classify the detected vehicles using convolutional neural networks according to several criteria. Thereafter I present the idea behind convolutional network based object detectors using Faster R-CNN. Here I introduce the development process of the network. Then I write about the installation, the used packages and the production of training data. Afterwards I discuss the fine tuning of the network according to our needs and introduce a possible architectural improvement. After producing right detections I try out a tracking algorithm.
The process just described has led to pleasing results, as I was able to create a non-motion based detector and integrate it with a well-functioning tracker.
The created system with the right improvements will be able to perform traffic counting even in complex environments. However the created detector can be used not only to solve traffic counting but also to find solutions to interesting problems.