In a vehicle, wheel speed sensors and inertial measurement units (IMUs) are present onboard, and their raw data can be used for localization estimation. Both wheel sensors and IMUs encounter challenges such as bias and measurement noise, which accumulate as errors over time. Even a slight inaccuracy or minor error can render the localization system unreliable and unusable in a matter of seconds. Traditional algorithms, such as the extended Kalman filter (EKF), have been applied for a long time in non-linear systems. These systems have white noise in both the system and in the estimation model. These approaches require deep knowledge of the non-linear noise characteristics of the sensors. On the other hand, as a subset of artificial intelligence (AI), neural network-based (NN) algorithms do not necessarily have these strict requirements. The current paper proposes an AI-based long short-term memory (LSTM) localization approach and evaluates its performance against the ground truth.

The advancement of autonomous vehicles (AVs) has brought forth a substantial need for effective education in robotic operating systems, particularly ROS 2, which serves as the backbone for many autonomous vehicle (AV) applications. This paper explores the academic approach and instructional methodologies tailored for teaching ROS 2 in the context of autonomous vehicle technology. It highlights the critical components and architecture of ROS 2, emphasizing its modularity, real-time communication capabilities, and robust ecosystem which make it ideal for AV development. Through a detailed curriculum outline, we describe hands-on learning activities, simulation-based exercises, and project-driven modules that facilitate deep understanding and practical skills acquisition. The effectiveness of these teaching methods is evaluated through a mixed-methods study involving student feedback, performance assessments, and project outcomes. Results indicate significant improvements in student comprehension and proficiency in both ROS 2 and autonomous vehicle systems. This research contributes to the body of knowledge by providing a comprehensive framework for educators to effectively teach ROS 2, thereby fostering the next generation of engineers proficient in developing and deploying autonomous vehicle technologies.

The prediction of terrain elevation values is a key task when it comes to off-road dynamics and inertial data estimation. A reliable elevation map can help in the estimation of future vehicle states and thus extend the response time window for autonomous navigation and control. We trained a deep learning model that is able to successfully predict top-down terrain depth maps in an off-road setting using a lightweight monocular depth estimation network. The labels were generated using a custom preprocessing algorithm to aid single image depth model training. Unlike other elevation estimation algorithms, our work can predict terrain variation from a higher camera setting without the use of a multi-sensor system. The network is also shown to work outside of the training data domain. The code will be available at https://www.github.com/norbertmarko/terrain-depth.

Inverse perspective mapping (IPM) is a crucial technique in camera-based autonomous driving, transforming the perspective view captured by the camera into a bird’s-eye view. This can be beneficial for accurate environmental perception, path planning, obstacle detection, and navigation. IPM faces challenges such as distortion and inaccuracies due to varying road inclinations and intrinsic camera properties. Herein, we revealed inaccuracies inherent in our current IPM approach so proper correction techniques can be applied later. We aimed to explore correction possibilities to enhance the accuracy of IPM and examine other methods that could be used as a benchmark or even a replacement, such as stereo vision and deep learning-based monocular depth estimation methods. With this work, we aimed to provide an analysis and direction for working with IPM.

Robust fusion of monocular and inertial data has the potential to offer a low-cost alternative for ground surface normal prediction ahead, compared to more expensive sensors, such as LiDAR. Yet robust camera-based prediction remains challenging, particularly for steep grades and texture-poor, homogeneous road surfaces. To address these issues, we propose an enhanced monocular camera-IMU fusion pipeline incorporating a lightweight transformer-based feature matcher for improved correspondence accuracy, and robust temporal filtering, using a spherical linear interpolation (SLERP) filter, to enhance consistency and reduce drift. To enable rigorous benchmarking and reproducibility, we also standardize the evaluation protocol and release a novel dataset containing synchronized camera, LiDAR, and IMU-derived pose data, specifically captured across diverse incline and decline scenarios. Extensive continuous validation demonstrates that our method significantly improves both accuracy and temporal stability over existing approaches, setting a new state of the art for robust, continuous ground normal estimation ahead. The proposed pipeline and dataset are publicly available at: https://norbertmarko.github.io/imu_cam_normal_prediction/ https://norbertmarko.github.io/imu_cam_normal_prediction/.