Z-ENG: 4-point polygon clustering and alignment with neural network

Félév: 2024-2025 ősz

Kategória: Szoftver

Téma leírása

One of the most important engineering tasks in the development of autonomous vehicles is to create a high-quality representation of the vehicle's real-world environment, as it is necessary for the robot to be able to make ideal decisions and navigate without external assistance. Although some details of the implementation may depend on the specific functionality, many general steps can be identified when considering such a solution. Such task is the detection, tracking and fusion of surrounding objects in time and space. These objects, which in real life often have complex shapes, can be described by simple geometric shapes, avoiding unnecessary overcomplication of algorithms and helping to achieve optimal runtimes. A common general shape is a rectangle, which can represent a parking space or even another vehicle.

It may also require the use of some prior knowledge, as adjacent parking spaces are usually parallel and oriented in the same direction. This can be used to further refine the detections, which are inherently noisy and imprecise. The validation of this heuristic can be done using a classic approach in two steps, first by clustering the objects that belong together (DBSCAN, K-means, mean shift, etc.) and then by aligning the resulting clusters together. However, with the appropriate training data, these two steps can be performed simultaneously using neural networks, substituting complex geometric algorithms, and allowing a higher level of generalization.

The student's task is to solve the clustering and alignment steps with neural networks for rectangle input: determine the optimal structure of the input, explore the neural network models suitable for this purpose, and then select the ideal candidate from them. It is also responsibility of the student to generate the training data (training-validation-test-set), implement the selected model, train, and evaluate it focusing on accuracy and runtime (relative comparison with a classical solution) in Python.

The student will be assisted by the Continental AI Development Center staff to solve the problem.

If you are interested in the topic, be sure to contact Dávid Sik by email before applying, indicating the selected topic, training level, major and the planned project subject.

Külső partner: Continental Autonomous Mobility Hungary

Maximális létszám: 1 fő