Deep learning based object detection on urban Lidar pointclouds

OData support
Dr. Harmati István
Department of Control Engineering and Information Technology

Artificial intelligence methods, especially the more and more popular deep learning can achieve significant results in many fields of science and technology. One of the most actively developed area of application is the self-driving car. It is very important for these cars to have algorithms capable of making or helping decision-making based on signals coming from different sensors. Several of the currently under development autonomous cars have a laser based sensor called LIDAR. The processing of the three-dimensional pointclouds generated by this device with artificial intelligence is an actively researched task.

Based on a publicly available research material I define my own way of solving the problem of pointcloud object classification. I specify and implement an algorithm that can convert the pointclouds to two-dimensional pictures, and I name this algorithm cylindrical projection. I create a framework built on a software library containing deep learning base functions which can train artificial intelligence models to classify pointcloud objects already transformed into pictures with cylindrical projection.

I define three models of different complexity, and I train these to perform the task at hand. I optimise the several parameter of the structure of these models and their training algorithms systematically. The result of several experiment for each model is an ideally set training algorithm and a file describing the best-performing instance of that model type.

In the case of all the trained models I analyse the average classification accuracy and the identification capability for every object class. Each of the models outperforms significantly the previous classical classification methods used for this task.


Please sign in to download the files of this thesis.