MTU Artificial Intelligence Colloquium

Presenter: Nirmal Loganathan, Data Science

Abstract: Precisely understanding the driving environment and determining the vehicle's accurate position is crucial for a safe automated maneuver. vehicle following systems that offer higher energy efficiency by precisely following a lead vehicle, the relative position of the ego vehicle to lane center is a key measure to a safe automated speed and steering control. This article presents a novel Enhanced Lane Detection technique with centimeter-level accuracy in estimating the vehicle offset from the lane center using the front-facing camera. Leveraging state-of-the-art computer vision models, the Enhanced Lane Detection technique utilizes YOLOv8 image segmentation, trained on a diverse world driving scenarios dataset, to detect the driving lane. To measure the vehicle lateral offset, our model introduces a novel calibration method using nine reference markers aligned with the vehicle perspective and converts the lane offset from image coordinates to world measurements. This design minimizes the sensitivity of offset estimation to lane detection accuracy and vehicle orientation. Compared to the existing deep learning-based depth perception models and stereo vision systems, our calibration method significantly improves postprocessing time and minimizes the impacts of the processing delay on the vehicle following system energy efficiency. To assess the accuracy and processing time, we implemented the model on an instrumented L4-capable vehicle and conducted automated vehicle following tests in a controlled environment. In our tests, the model achieved a high level of accuracy, with a biased error of only 0.214 m and a random walk error standard deviation of 0.135 m, demonstrating its reliability across various environmental conditions and ensuring precise lane tracking. Results demonstrate reliable performance across various environmental conditions and sensor noise levels, ensuring precise lane tracking and enhanced automated maneuvering.

Presenter: Kevin Li, Computer Science

Abstract: Accurate mapping is crucial for the safe operation of Uncrewed Surface Vessels (USV), especially when relying on 3D LiDAR sensors, which are susceptible to noise from factors like water spray and turbulent surface conditions. This paper investigates the performance of two prominent 3D LiDAR mapping algorithms—HDL graph SLAM and Point LIO—when subjected to injected noise models that mimic real-world maritime disturbances. By introducing controlled noise into the sensor data, we evaluate how each algorithm maintains mapping accuracy under degraded conditions. Quantitative metrics such as Structural Similarity Index Measure (SSIM) and Intersection over Union (IoU) are utilized to assess and compare the robustness of the generated maps. The study will highlight the open-source ROS mapping algorithms strengths and limitations in noisy environments. These findings aim to inform the development of more robust mapping strategies, ultimately enhancing the reliability and safety of USV navigation in challenging maritime settings.

Presenter: Tagore Kosireddy, Computer Science

Abstract: Autonomous drone landing presents significant challenges due to complex aerodynamic dynamics and real-time control requirements. Our work implements a Deep Reinforcement Learning (DRL) framework for autonomous drone landing using PyBullet physics real time simulations. We are using modular architecture integrating PID control with state-of-the-art DRL algorithms (PPO, SAC, TRPO, TD3, A2C, etc) and benchmarking various Multiple reward functions (exponential, temporal difference, orientation-aware) derived from quadrotor dynamics research. The environment features 12D observation spaces (position, velocity, orientation) and 3D continuous action spaces for thrust vector control. A novel curriculum learning approach progressively increases landing difficulty through randomized initial positions and adaptive waypoint generation. The implementation leverages parallelized training across CPU cores and includes safety constraints for real-world deployment considerations.

Presenter: Nazanin Mahjourian (Mechanical and Aerospace Engineering)

Abstract: The success of AI in industrial applications is heavily dependent on the quality of the datasets used to train models. However, large-scale datasets often suffer from label noise, inconsistencies, and errors, which can negatively impact model performance. This problem is particularly pronounced in industrial domains, where obtaining high-quality labels is costly and time-consuming. In this presentation, we discuss a method named LabelAlchemy to sanitize the labels of a large industrial dataset, FactoryNet, which contains both human-generated and web-scraped labels using CLIP vision-language model. The CLIP model is a state-of-the-art vision-language model that has a dual-encoder architecture with separate encoders for images and text based on transformers, which employ multi-head self-attention mechanisms to model complex relationships within sequences. The text encoder processes tokenized input sequences and captures contextual nuances of words within a sentence. Similarly, the image encoder processes visual features extracted from input images. The resulting embeddings are mapped to a common latent space, allowing CLIP to measure the similarity between image-text pairs effectively. LabelAlchemy is shown to sanitize the FactoryNet dataset and improve the performance of classification algorithms.