Optical flow algorithms in driver assistance systems

OData support
Supervisor:
Dr. Blázovics László
Department of Automation and Applied Informatics

Optical flow is an image processing method, which is capable of describing tangential movements through sequence of images. Therefore, this description is an information which is becoming more sought in the driver assistance systems. Considering the pace of the automation of the vehicles the optical flow will be nearly invaluable in the future.

The goal of my thesis is to investigate the possible computation methods of the optical flow. I have compared different implementations in regards of the computation time and precision, which will be presented in this thesis. Another consideration in my comparisons was the possibility of using a single frontal camera in the vehicle, along with the hardware requirements. Lastly, I have analyzed the possible usages of the optical flow in this environment.

I have divided the possible algorithms into three groups. For all these groups I have investigated the possible precision and runtime. To compare these algorithm I had to choose a benchmark with which I was able to give a percentage as precision. In my thesis, I have selected the KITTI Vision Benchmark for the evaluation.

The three selected approaches were the following: Sparse Optical Flow, Dense Optical Flow and Deep Learning. For each method I, have selected an algorithm to represent the general precision and computation time of the group. My tests concluded that the event though the Sparse Optical Flow was precise and fast it was not able to calculate a large number of vectors, therefore its usage is limited to the area where the pixels haven been calculated. The Dense Optical Flow algorithms are slow but capable to reach a higher precision. Lastly, the Deep Learning algorithms were both fast and precise if a large amount of training data can be provided. In my work, I have presented two examples for the generation of the training data with the combination of Dense Optical Flow and semantic segmentation algorithms.

Ultimately, I have tested the chosen deep learning algorithm in three different hardware and tested it in a real vehicle environment. In this test, I have concluded the precision in percentages in regard of the vehicle’s speed, furthermore I have presented some interesting use cases for the Optical Flow.

In the course of my work, I tried to achieve progressive results and keep in mind the advancements of the technology with which new aspects of the Optical Flow can be exploited. In regards of these aspirations, I was able to attend two international conferences: Max Planck Lecture 2017 in Stuttgart (Mobileye) and GPU Technology Conference Europe 2017 in Münich (Nvidia). At these conferences, I was able to get a deeper look into the future of optical flow and the newest advancements.

Downloads

Please sign in to download the files of this thesis.