Robust multispectral image-based localisation solutions for autonomous systems

With the recent increase of interest in multispectral imaging, new image-based localisation solutions have emerged. However, its application to visual odometry remains overlooked. Most localisation techniques are still being developed with visible cameras only, because the portability they can offer and the wide variety of cameras available. Yet, other modalities have great potentials for navigation purposes. Infrared imaging for example, provides different information about the scene and is already used to enhance visible images. This is especially the case of far-infrared cameras which can produce images at night and see hot objects like other cars, animals or pedestrians. Therefore, the aim of this thesis is to tackle the lack of research in multispectral localisation and to explore new ways of performing visual odometry accurately with visible and thermal images. First, a new calibration pattern made of LED lights is presented in Chapter 3. Emitting both visible and thermal radiations, it can easily be seen by infrared and visible cameras. Due to its peculiar shape, the whole pattern can be moved around the cameras and automatically identified in the different images recorded. Monocular and stereo calibration are then performed to precisely estimate the camera parameters. Then, a multispectral monocular visual odometry algorithm is proposed in Chapter 4. This generic technique is able to operate in infrared and visible modalities, regardless of the nature of the images. Incoming images are processed at a high frame rate based on a 2D-to-2D unscaled motion estimation method. However, specific keyframes are carefully selected to avoid degenerate cases and a bundle adjustment optimisation is performed on a sliding window to refine the initial estimation. The advantage of visible-thermal odometry is shown on a scenario with extreme illumination conditions, where the limitation of each modality is reached. The simultaneous combination of visible and thermal images for visual odometry is also explored. In Chapter 5, two feature matching techniques are presented and tested in a multispectral stereo visual odometry framework. One method matches features between stereo pairs independently while the other estimates unscaled motion first, before matching the features altogether. Even though these techniques require more processing power to overcome the dissimilarities between V multimodal images, they have the benefit of estimating scaled transformations. Finally, the camera pose estimates obtained with multispectral stereo odometry are fused with inertial data to create a robustified localisation solution which is detailed in Chapter 6. The full state of the system is estimated, including position, velocity, orientation and IMU biases. It is shown that multispectral visual odometry can correct drifting IMU measurements effectively. Furthermore, it is demonstrated that such multi-sensors setups can be beneficial in challenging situations where features cannot be extracted or tracked. In that case, inertial data can be integrated to provide a state estimate while visual odometry cannot.