Radim Šára

Aims. We applied machine learning to the entire data history of ESO’s High Accuracy Radial Velocity Planet Searcher (HARPS) instrument. Our primary goal was to recover the physical properties of the observed objects, with a secondary emphasis on simulating spectra. We systematically investigated the impact of various factors on the accuracy and fidelity of the results, including the use of simulated data, the effect of varying amounts of real training data, network architectures, and learning paradigms. Methods. Our approach integrates supervised and unsupervised learning techniques within autoencoder frameworks. Our methodology leverages an existing simulation model that utilizes a library of existing stellar spectra in which the emerging flux is computed from first principles rooted in physics and a HARPS instrument model to generate simulated spectra comparable to observational data. We trained standard and variational autoencoders on HARPS data to predict spectral parameters and generate spectra. Convolutional and residual architectures were compared, and we decomposed autoencoders in order to assess component impacts. Results. Our models excel at predicting spectral parameters and compressing real spectra, and they achieved a mean prediction error of ~50 K for effective temperatures, making them relevant for most astrophysical applications. Furthermore, the models predict metallicity ([M/H]) and surface gravity (log g) with an accuracy of ~0.03 dex and ~0.04 dex, respectively, underscoring their broad applicability in astrophysical research. Moreover, the models can generate new spectra that closely mimic actual observations, enriching traditional simulation techniques. Our variational autoencoder-based models achieve short processing times: 779.6 ms on a CPU and 3.97 ms on a GPU. These results demonstrate the benefits of integrating high-quality data with advanced model architectures, as it significantly enhances the scope and accuracy of spectroscopic analysis. With an accuracy comparable to the best classical analysis method but requiring a fraction of the computation time, our methods are particularly suitable for high-throughput observations such as massive spectroscopic surveys and large archival studies.

Cameras are the prevalent sensors used for perception in autonomous robotic systems, but their initial calibration may degrade over time due to dynamic factors. This may lead to a failure of downstream tasks, such as simultaneous localization and mapping (SLAM) or object recognition. Hence, a computationally lightweight process that detects the decalibration is of interest. We describe a modification of StOCaMo, an online calibration monitoring procedure for a stereoscopic system. The method uses robust kernel correlation based on epipolar constraints; it validates extrinsic calibration parameters on a single frame with no temporal tracking. In this paper, we present a modified StOCaMo with an improved recall rate on small decalibrations through a confirmation technique based on resampled variance. With fixed parameters learned on a realistic synthetic dataset from CARLA, StOCaMo and its proposed modification were tested on multiple sequences from two real-world datasets: KITTI and EuRoC MAV. The modification improved the recall of StOCaMo by 25 % (to 91 % and 82 %, respectively), and the accuracy by 12 % (to 94.7 % and 87.5 %, respectively), while labeling at most one-third of the input data as uninformative. The upgraded method achieved the rank correlation between StOCaMo V-index and downstream SLAM error of 0.78 (Spearman).

The relative poses of visual perception sensors distributed over a vehicle's body may vary due to dynamic forces, thermal dilations, or minor accidents. This paper proposes two methods, OCAMO and LTO, that monitor and track the LiDAR-Camera extrinsic calibration parameters online. Calibration monitoring provides a certificate for reference calibration parameters validity. Tracking follows the calibration parameters drift in time. OCAMO is based on an adaptive online stochastic optimization with a memory of past evolution. LTO uses a fixed-grid search for the optimal parameters per frame and without memory. Both methods use low-level point-like features, a robust kernel-based loss function, and work with a small memory footprint and computational overhead. Both include a preselection of informative data that limits their divergence. The statistical accuracy of both calibration monitoring methods is over 98%, whereas OCAMO monitoring can detect small decalibrations better, and LTO monitoring reacts faster on abrupt decalibrations. The tracking variants of both methods follow random calibration drift with an accuracy of about 0.03° in the yaw angle.

Certain classification tasks in computer vision require the classifier response to be computed in every pixel of an image. When combined with large, complex features, it becomes challenging to build such a classifier on a standard PC architecture and achieve real-time performance. We present an FPGA implementation of a car wheel classifier response computation, built as an instantiation of a generic classification system. An interesting optimization problem concerning performance and speed is addressed. Our implementation is running in real-time as a part of a more complex collision mitigation system based on car detection in video data.

Current collision mitigation systems focus on rear end collisions. To address the full spectrum of real world accidents, these systems will have to be enhanced to cover more traffic situations. Vehicle to vehicle accidents in crossing traffic situations make up around 25% of accidents in Germany. This paper discusses the requirements and differences compared to rear-end collisions. Presented here is an action concept that takes into account how the impact configuration is changed by breaking the host (impacting) vehicle. Based on this concept the requirements for the detection of crossing traffic were derived. These requirements were met by developing a video system based on a monocular wide field of view camera. It is further shown how this action concept and sensor were integrated into a demonstrator vehicle and evaluated in full scale testing.

We present a method for calculation of disparity maps from stereo sequences. Disparity map from previous frame is first transferred to the new frame using estimated motion of the calibrated stereo rig. The predicted disparities are validated for the new frame and areas where prediction failed are matched with a traditional stereo matching algorithm. This method produces very fast and temporally stable stereo matching suitable for real-time applications even on non-parallel hardware.