Vehicle detection and tracking is a long-researched area. The algorithms are constantly evolving, but a general methodology still does not exist for the solution. During the development of transport infrastructure the statistical analysis of vehicle traffic is crucial and the well-founded development around the world allows for savings of billions of euros each year. The aim is to build a system which sufficiently accurately and efficiently solves the detection and tracking of vehicles by using image processing techniques.
For the realization I used deep learning. Initially, I dealt a lot with data production and augmentation to train AlexNet classifier network. While I was reseacrhing for SSD , I developed the workflow of data production and programs for data processing and I also installed some programs for it. I built up the necessary environment for SSD and I installed the required libraries. I completed trainings with SSD network with 300x300 and 500x500 inputs, taking into account the goals and parameters of the network. Using the SSD output, I connected the SORT named tracker with the detector. With the cooperation between the detector network and tracker I produced acceptable quality of results in regard to accuracy and speed.
By further developing the detector network and tracker, we will be able to produce accurate and fast results for special circumstances. The detector and classifier network can be applicable to many other problems by re-training without architecture modification