This dissertation presents several perception technology for autonomous driving in the context of BMW's HAD(Highly Automatic Driving) cars.

To begin with, the combined use of Multi-LRF with cameras has been increasingly popular in self-driving automobile. In order to convert multiple LRF data into a unified coordinate system, we have to obtain the rigid transformation among multi-LRF. Hence, we propose a new algorithm for online extrinsic calibration of multi-LRFs by observing a planar checkerboard pattern and solving for transformation between the views of a planar checkerboard from a camera and multi-LRF. Compared with traditional algorithm, our algorithm outperform in two ways. Firstly, adopting the noise of images and LRF depth readings, we can exactly calculate the exa...[ Read more ]

This dissertation presents several perception technology for autonomous driving in the context of BMW's HAD(Highly Automatic Driving) cars.

To begin with, the combined use of Multi-LRF with cameras has been increasingly popular in self-driving automobile. In order to convert multiple LRF data into a unified coordinate system, we have to obtain the rigid transformation among multi-LRF. Hence, we propose a new algorithm for online extrinsic calibration of multi-LRFs by observing a planar checkerboard pattern and solving for transformation between the views of a planar checkerboard from a camera and multi-LRF. Compared with traditional algorithm, our algorithm outperform in two ways. Firstly, adopting the noise of images and LRF depth readings, we can exactly calculate the exact position and pose of the checkerboard that can largely reduce the transformation error. Secondly, the completion calibration process is online, which means the exact position and pose of the checkerboard can be obtained in real-time and manipulated by robotic arm.

A second key challenge in autonomous driving is scene understanding. One of fundamental basis of scene understanding is scene labeling. Its objective is to recognize which scene category each pixel describes. Using calibrated sensors, we can obtain high-resolution images as the input of our deep learning architecture. In terms of feature extraction, it has been proved in this context that deep learning architecture outperform traditional algorithms. We set up an open-source c++ library, namely libcnn++, for scene labeling. Our library focuses on Convolutional Neural Network for feature extraction and description. Functions of pre-processing for the raw input images, like space transformation and normalization , are also included in this library. Existing deep learning libraries, such as Caffe and Theano, could be easily adapted to object and face recognition. However, they demand much additional tough work for pixel-wise recognition. In contrast to these libraries, our proposed library is much easier to handle, with emphasis on robotic applications.