Estimation of Lidar Point Clouds Based on Ultrasonic Sensors Using Deep Learning
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Complex adaptive systems (MPCAS), MSc
Engineering mathematics and computational science (MPENM), MSc
Engineering mathematics and computational science (MPENM), MSc
Publicerad
2023
Författare
Hjalmarsson, Carl
Jäghagen, Jesper
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Safety is a key point of research in the automotive industry. Car companies dedicate
a great amount of time and resources to make their cars safer by developing sensory
systems like Low Speed Perception (LSP). In this thesis, we have explored the possibility to enhance the contribution of ultrasonic sensors to LSP by leveraging deep
learning to mimic data from the more expensive Lidar sensors.
To do this we draw inspiration from three different deep learning approaches: image denoising, image segmentation and image-to-image translation, resulting in five
different models: DAE(bin), DAE(mse), UNET(bin), UNET(ce) and an I2I cGAN. We
train these models on three datasets, each containing paired ultrasonic and Lidar
representations of a two-dimensional environment around the car. In order to measure the performance of these models, we develop an evaluation framework where we
assess the ability of the models to map ultrasonic object detections to corresponding
Lidar detections.
We find that the performance of all networks is highly dependent on the data representation. When using a basic representation, consisting only of point detections
and free-space, all models fail to improve upon the baseline ultrasonic sensor score.
When using a sparse representation consisting of detection arcs, however, UNET(bin)
succeeds in outperforming the baseline and mimicking the more accurate Lidar representation (see cover). Finally, when adding unknown areas of the environment to
the sparse representation, both UNET(ce) and the cGAN manage to outperform the
baseline in most aspects, and we see a convergence towards the more realistic Lidar
representation.
The results show that there is indeed a possibility to enhance ultrasonic sensor perception using deep learning and Lidar reference data, and while there is still much
room for improvement, we have shown that there is potential in further research on
this task.
Beskrivning
Ämne/nyckelord
Low Speed Perception, Ultrasonic Sensor Perception, Deep Learning, Autoencoders, U-Net, Conditional Generative Adversarial Networks.