Excel to HTML
MONDAY 2 MARCH 2026
Perception, Mapping, and Prediction Algorithms
Session Chair: Patrick Denny, University of Limerick
15:30 - 17:30
Grand Peninsula G
15:30AVM-100
A comparative analysis of video- and pose-based action recognition for in-cabin driver monitoring, Lukas Brunner, Tu Wien (Austria); Dominik Schoerkhuber, TU Wien (Austria) [view abstract]
We present a comparative study of pose-based vs. video-based Human Action Recognition (HAR) methods for driver monitoring in car cockpits. In this context, comparisons of neural network architectures from the field of deep learning-based video understanding are scarce. However, pose- and video-based HAR has significant potential for advanced driver-assistance systems in semi-autonomous driving on public roads. We compare prediction performance, per-class false-negative rate, model size, computational requirements, and inference latency on the established Drive&Act and the proprietary Driver Action Insight datasets. While the diversity and scale of available datasets make comparisons challenging, results suggest that both approaches benefit from pretraining, but pose- and video-based techniques perform differently for specific action classes, such as those that depend on body motion or the appearance of objects.
15:50AVM-101
Real-time online learning trajectory prediction via efficient latent predictor, Jierui Peng, Case Western Reserve University (US); Yu Yin, Case Western Reserve University (US); Vipin Chaudhary, Case Western Reserve University (US) [view abstract]
Trajectory prediction is crucial for autonomous systems, but traditional deep learning models, typically trained on specific pre-collected trajectories, often fail to generalize to unseen scenarios due to distribution shifts. Recent approaches address this by integrating online learning for adaptive deployment. However, existing online learning methods face two major challenges: (1) long training times, which prevent real-time execution, and (2) failure to account for variations in input data speed, leading to performance degradation when processing high-speed dynamic scenarios. To overcome these limitations, we introduce a latent-space predictor that forecasts future trajectories by aligning learned latent representations with encoded ground truth. This approach enhances robustness to distribution shifts while reducing reliance on direct coordinate regression. Additionally, we incorporate a lightweight online learning module, enabling efficient real-time adaptation without full model retraining. We evaluate our method on nuScenes, Waymo, and Lyft L5 datasets, focusing on data distribution shift scenarios. Experimental results demonstrate that our model outperforms state-of-the-art online learning methods, achieving approximate 9.9% improvement in trajectory prediction accuracy while significantly reducing optimization time up to 54%.
16:10AVM-102
A semi-decentralized collaborative framework for lightweight map updates in urban navigation systems, Gaethan Kevin MOUNGOUE NJIYEP, Institut VEDECOM (France); Imeen Ben Salah, Institut VEDECOM (France); Ahmed Rafik Islem BELHADEF, Institut VEDECOM (France) [view abstract]
Autonomous vehicles currently rely on High-Definition (HD) maps for precise localization and path planning. However, traditional HD mapping approaches suffer from high costs, inherent rigidity, and slow update cycles, making them inadequate for dynamic urban environments. This paper presents a novel lightweight collaborative mapping architecture that enables real-time map updates through multi-agent cooperation. Our approach combines Joint Compatibility Branch and Bound (JCBB) for data association, Dempster-Shafer Theory (DST) for uncertainty quantification and landmark classification, and Extended Kalman Filter (EKF) for landmark pose estimation. Experimental validation using the CARLA simulator demonstrates accurate landmark classification and localization. Furthermore, collaborative data fusion reduces false positives and improves overall system reliability.
16:30AVM-103
Automatic registration of subsea LiDAR point clouds, Josephine Clapp, Student at Rochester Institute of Technology (US); Carl Salvaggio, Professor at Rochester Institute of Technology, Advisor (US); Byron Eng, Rochester Institute of Technology [view abstract]
Subsea LiDAR technology is increasingly used in the oil and gas industry for asset inspection and monitoring, where accurate 3D mapping supports structural integrity assessment, maintenance planning, and environmental monitoring. A key challenge lies in the registration of multiple adjacent/overlapping LiDAR scans. Current workflows rely on manual alignment, which is timeconsuming and prone to error. This project investigates the use of the Iterative Closest Point (ICP) algorithm, implemented through the Open3D library, to automate subsea LiDAR point cloud registration. Using datasets of subsea infrastructure and seabed, the workflow was benchmarked against manual methods that typically achieve alignment accuracy of <5 mm. Preliminary results show that the automated process achieves comparable accuracy while significantly reducing processing time and human intervention. This project shows that replacing manual alignment with an automated solution is possible, lowering operational costs and improving reliability.
TUESDAY 3 MARCH 2026
Advances in Sensing Technologies for Automotive
Session Chair: Patrick Denny, University of Limerick
11:00 - 12:20
Grand Peninsula G
11:00AVM-107
AVM KEYNOTE: Introduction to the latest mmWave automotive imaging radar technologies, [view abstract]
In this talk, an overview of the latest automotive radar technologies will be provided. Topics include the latest system architecture, high-resolution imaging radar processing, massive MIMO waveform design, and radar-to-radar interference mitigation concepts. Field test results of a state-of-the-art 24 Tx-by-24 Rx imaging radar achieving lidar-like point clouds in challenging environments will also be presented in the end
12:00AVM-108
Low power automotive vision using hybrid sensors on NPUs, Kamal Rana, Omnivision Technologies Inc (US); Shijie Xiao, Omnivision Technologies Inc (US); Zhongyang Huang, Omnivision Technologies Inc (US); Bo Mu, Omnivision Technologies Inc (US) [view abstract]
Automotive vision is a key component of advanced driver assistance systems (ADAS), enhancing road safety and improving vehicle operation for drivers. A critical requirement for automotive vision is achieving faster detections to ensure higher safety. However, faster object detections using intensity-based CMOS Image Sensors (CIS) is limited by their frame rate. While increasing the intensity-based CIS frame rate enables faster object detection, it also results in higher sensor data rates and significantly increasing power consumption. In our previous work, we demonstrated that utilizing event-based pixels, which provide sparse but high-temporal-resolution data, with low intensity-based CIS framerate offers an alternative solution for faster object detections for automotive vision. Using hybrid sensors data (low intensity-based CIS framerate + event pixels data ) achieves comparable performance as high intensity-based CIS framerate but with reduced data rates and power consumption. For example, in our previous study, we showed that using 7 fps intensity-based CIS data combined with event pixels data delivers the same performance as 20 fps intensity-based CIS data, but with 40% lower data rate. In this work, we are implementing post-training quantization and quantization aware training techniques to automotive vision models trained on hybrid sensor data (intensity-based CIS framerate + event pixels data). This enables automotive vision models using hybrid sensors to reduce both sensor data rates and power consumption during inference, particularly when deployed on Neural Processing Units (NPUs).
12:20AVM-109
High sensitivity automotive color filter arrays with improved color separation and accuracy, Tripurari Singh, Image Algorithmics (US); Mritunjay Singh, Image Algorithmics (US); Arnav Singh, (US) [view abstract]
The automotive industry has developed several CFAs by substituting the green and blue filters in standard RGB Bayer arrays with lighter colors, while retaining red due to the importance of red lights and signs. This strategy is sub-optimal, as the critical red-yellow traffic light separation relies on the accurate spectral response of the green filter, not red. Furthermore, red color filters have spectral responses that compromise color accuracy for noise, and red is more accurately captured as a difference between yellow and green signals.Based on these observations, we propose the GYYCy CFA, which provides better color separation of traffic features, improved color accuracy, and enhanced dynamic range compared to the incumbent RCCB, RCCG, RYYCy, and RYYB CFAs. The sensitivity of GYYCy exceeds that of RYYCy and RYYB and remains competitive with RCCB and RCCG. Like other CFAs popular in automotive applications, GYYCy is compatible with the standard Bayer demosaicking algorithms.Moving beyond the traditional 2x2 Bayer layout, we further propose a 4x4 YGB CFA that achieves better color accuracy and separation of red-yellow traffic lights than RGB Bayer, while maintaining the sensitivity and dynamic range of GYYCy. Replacing Y, G, B with L, M, S respectively further enhances sensitivity and results in a colorimetric camera.
Physically Accurate Simulation in Automotive Imaging
Session Chair: Patrick Denny, University of Limerick
09:30 - 10:30
Grand Peninsula G
09:30AVM-104
CIDPL: A real-time CUDA-accelerated Python-framework simulating PSF-based optical artifacts integrated in MMDetection, Maximilian Dornik, ; Julian Barthel, ; Daniel Jakab, ; Alexander Braun, [view abstract]
Object detection is central to autonomous driving, where reliability is vital for collision avoidance. While frameworks such as PyTorch or MMDetection enable fast and flexible evaluation of AI models, they lack tools to systematically assess the impact of point spread function (PSF)-based optical artifacts, which impair obstacle recognition and pose safety risks.We present the first integration procedure of a physically accurate, PSF-based optical artifact simulation pipeline into MMDetection. Our approach adapts the CUDA-accelerated Image Degradation Library (IDL), developed in our working group, which supports position-dependent degradation using PSF grids and is validated through Modulation Transfer Function (MTF)-curves. To bridge IDL with MMDetection, we introduce the CUDA Image Degradation Python Library (CIDPL), enabling automated, real-time evaluation without pre-processing or storing degraded datasets.Our approach addresses key bottlenecks: minimizing CPU-GPU transfers, maintaining high throughput, reducing evaluation overhead, and enabling batch parallelism. Beyond test-time augmentation (TTA), our approach preserves back-traceability between degraded variants and evaluation results. To this end, we propose data structures with variation re-assignability, supporting multiple perturbations per image while fully exploiting data parallelism under memory constraints.Initial experiments demonstrate substantial speedups and reduced configuration complexity, contributing to more robust AI models evaluated under realistic optical conditions, thereby enhancing safety.
09:50AVM-105
ISET-LFM: A physics-based simulation framework and dataset for LED flicker in automotive imaging, Ayush Jamdar, University of California San Diego (US) [view abstract]
LED flicker is a persistent artifact in imaging, where lights modulated via Pulse Width Modulation (PWM) above 90 Hz appear steady to humans but produce temporal intensity variations in captured video. While hardware mitigations like split-pixel architectures reduce flicker, they introduce a fundamental trade-off with motion blur. Progress in learned LED flicker mitigation (LFM) is currently hindered by a lack of public ground-truth datasets. We address this gap with ISET-LFM, an open-source physics-based simulation framework that models LED flicker in driving scenes. Built on the ISET ecosystem, our pipeline combines camera motion simulation with an analytical flicker model to generate realistic dual-exposure frame sequences alongside flicker-free ground truth. We provide a synthetic dataset of scene radiance, enabling benchmarking and training of LFM algorithms across diverse sensor and ISP architectures.
10:10AVM-106
The influence of Monte Carlo denoisers on the quality of real-time optical system simulations, Jan Halama, University of Applied Sciences D sseldorf (Germany); Alexander Braun, University of Applied Sciences D sseldorf (Germany) [view abstract]
Raytracing in combination with Monte Carlo simulation is an accurate method to simulate optical systems in virtual 3D scenes. Since Monte Carlo simulation relies on random sampling, many samples per pixel need to be computed for a noise-free image, resulting in high computational effort. Even the fastest ray tracers can only trace a few samples per pixel in real-time. A common solution in computer graphics is to compute the image with a few samples per pixel and apply a Monte Carlo denoiser to remove the noise. Since the denoiser alters the image, the question arises to what extent this influences the quality of the simulation. Utilizing the Simulating tests to test simulation -method, we measure the MTF function of a simulation denoised with the NVIDIA OptiX Denoiser and compare it with a highly sampled baseline simulation. We investigate if although the image is altered by denoising, using denoised ray tracing simulations yields more realistic results for real-time rendering than other non-physically based methods such as gaussian blurs. Beside real-time applications, we implemented our novel real-time capable Ray-LUT approach into PBRT and investigate potential time savings for offline rendering.
Special Session on Standards and Metrics for Automotive Perception
Session Chair: Patrick Denny, University of Limerick
15:30 - 17:30
Grand Peninsula G
15:30AVM-110
Update IEEE P2020 automotive system image quality standard, Shaheen Amanullah, OnSemi [view abstract]
The IEEE P2020 working group published Revision 1 of the Automotive System Image Quality Standard in early 2025. To address rapid advancements in imaging technologies for ADAS and autonomous driving, a new Project Authorization Request (PAR) has initiated Revision 2, targeted for completion by end of 2028. This update focuses on refining key performance indicators (KPIs) to ensure relevance for modern automotive imaging systems. Core areas of development include: Convergence standardizing stabilization of algorithms under dynamic lighting; Color enhancing methodology to measure color separation and fidelity for both human and machine vision; Noise Covers metrics and procedures for operation in HDR mode and test time efficiency; Noise Extension Includes detailed separation of fixed-pattern noise components and adaption of the noise equivalent quanta metric to automotive image sensors and cameras; SFR Extension moving from a best case scenario (lab environment) to a real case scenario (motion in camera and object, impact of pixel design on SFR) to better reflect the performance when deployed; and Compression establishing objective methods to quantify compression artifacts and their impact on computer vision integrity. These efforts aim to maintain a robust, forward-looking standard that supports safety-critical imaging in next-generation vehicles.
15:50AVM-111
Recommendations for road markings to improve machine vision in autonomous driving, Brian Deegan, University of Galway (Ireland); Robin Jenkin, NVIDIA (US) [view abstract]
Road markings have been developed and standardized for human perception for over a century. With the increasing presence of autonomous and semi-autonomous vehicles, which rely heavily on machine vision, new challenges emerge for road marking design. Current markings optimized for human drivers may not provide optimal performance for automated perception systems, particularly under diverse environmental and lighting conditions. This paper addresses the need to reassess marking standards in the context of mixed traffic environments. We identify key factors influencing detectability by both humans and machines, including color, contrast, retroreflectivity, and spatial consistency. The discussion highlights potential conflicts between human-centric and machine-centric requirements and considers how design choices may affect safety, robustness, and interoperability. Based on this analysis, we propose recommendations for evolving road marking standards that support reliable machine vision while maintaining legibility and safety for human road users, and provide updates from the IEEE P2020 Automotive Image Quality Standards working group.
16:10AVM-112
Mahalanobis distance as a robust metric for color separation, Robin Jenkin, NVIDIA (US) [view abstract]
The discrimination of colors is a fundamental requirement for autonomous driving, where the ability to detect differences between traffic signals, road markings, and emergency vehicle lights is paramount for safety. Traditional methods for quantifying color separation, such as ?E or ?C are insufficient as they fail to account for sensor noise, which is further significantly altered by image signal processing. Additionally, Euclidian distance-like measures are not invariant to linear transforms, such as those encountered when applying a color correction matrix or white balance and thus pre- and post-CCM metrics can behave very differently. This paper proposes the use of the Mahalanobis distance as a statistically robust metric for evaluating color separation, as it explicitly incorporates the covariance of color channel noise.The color filter array (CFA) and subsequent color correction matrix (CCM) are the primary determinants of an imaging sensor's color fidelity and noise characteristics. More transparent CFA patterns such as RCCB and RCCG enhance sensitivity at the cost of increased cross-correlation of noise between color channels during color correction which amplifies noise, and distorts noise clusters into oriented, elliptical clouds. When this occurs, Euclidean distance becomes a misleading indicator of separability, as a simple gain transformation can artificially increase or decrease the distance without genuinely improving the signal-to-noise ratio of the color difference. The change in the Euclidian distance is dependent on the dot product between the difference vector between the colors and the axis of the maximum gain of the CCM.Using a radiometric simulation of image sensors with various CFAs, the imaging of a color checker chart with various illumination sources, the propagation of noise is examined to calculate various pre- and post-CCM color separation metrics. By analyzing the Mahalanobis distance for critical color pairs, it is shown that this metric provides an invariant measure of separation that correctly accounts for the shape, size, and orientation of the noise distributions. The effects of non-linear transformations, such as conversion to CIELAB and gamma correction on the Mahalanobis distance are further examined. It is shown that Mahalanobis distance provides a method for robustly estimating the differences of color signals entering the ISP that is immune to the effects of CCM and AWB.
Panel: Developing Road Markings Standards for Machine Vision in ADAS, [view abstract]
Panel discussion moderatorPatrick Denny, University of LimerickParticipantsRobert Dingess, president, Traffic Safety Research InstituteDavid Entrekin, KTA-Tator & FutureLabsBrian Deegan, University of GalwayPaul Romanczyk, Imatest
WEDNESDAY 4 MARCH 2026
Automotive Imaging Performance I (Joint Session with Image Quality and System Performance)
Session Chair: Elaine Jin
15:30 - 16:30
Grand Peninsula G
15:30AVM-113
Balancing exposure vs. resolution in high-speed ADAS imaging, Gabriel Bowers, Mobileye (France); Uwe Artmann, Image Engineering (Germany); Max Gade, Image Engineering (Germany) [view abstract]
The automotive industry has made significant strides in improving camera technologies for ADAS and autonomous vehicles, with a strong focus on increasing resolution, achieving flicker-free dynamic range, and improving low light sensitivity. These improvements are challenged when driving at increasing speeds, as shorter exposures are needed to detect nearby objects at the same resolution. A representation of such effects is often missing from the camera IQ tests. However, such analysis can be valuable as it can better clarify the system design trade-off between caSummary Imagemera specifications (e.g., resolution, lens aperture) and exposure control design. This paper will provide an attempt to highlight how this analysis of a system implementation can be achieved in lab settings.
15:50AVM-114
brilliantISP: An open source HDR ISP for research, Brian Deegan, University of Galway (Ireland) [view abstract]
BrilliantISP: An Enhanced HDR Image Signal Processing Pipeline Abstract This paper presents BrilliantISP, an enhanced high dynamic range (HDR) image signal processing pipeline designed to advance computational photography capabilities. Building upon the foundation of infiniteISP, FastOpenISP, and OpenISP, our implementation introduces several key innovations to address limitations in existing open-source ISP solutions. The enhanced pipeline incorporates a novel decompanding function that effectively linearizes companded sensor data, enabling more accurate downstream processing. We have integrated Durand's HDR tone mapping algorithm, providing superior dynamic range compression while preserving local contrast and detail visibility. The system features modified bit depth handling throughout the pipeline to maintain precision during HDR processing operations. Significant performance optimizations have been achieved through algorithmic improvements and execution time optimization, with comprehensive debug logging capabilities for development and research applications. The modular architecture supports flexible configuration through YAML parameter files, enabling rapid prototyping and experimentation. Current development focuses on implementing HDR multicapture merge functionality and lens shading correction algorithms. Preliminary results demonstrate improved image quality metrics compared to baseline implementations, particularly in high contrast scenes. The open-source nature of BrilliantISP facilitates reproducible research and collaborative development in computational photography. This work contributes to the growing ecosystem of open-source ISP solutions, providing researchers and developers with enhanced tools for HDR image processing applications in both academic and commercial contexts.
16:10AVM-115
Information-based dynamic range, [view abstract]
We present a new approach to measuring camera dynamic range and low-light performance based on C4 information capacity, which is measured directly from ISO 12233-standard 4:1 contrast slanted edges. Our initial technique involves photographing a test chart that contains 4:1 slanted edges over an extremely wide range of exposures, from or second (where the brighter side of the edge saturates) to 1/2000 or 1/4000 second, where the image appears nearly black, but a noisy edge is still present. The major advantages of this method are1. Dynamic range limits are based on an actual performance metric (C4) rather than SNR, which is only one of the factors that contributes to camera performance.2. C4 correctly handles performance degradation due to stray light.In the final paper we will discuss new techniques, still under development, for facilitating the measurement.
Automotive Imaging Performance II (Joint Session with Image Quality and System Performance)
16:30 - 17:30
Session Chair: Elaine Jin
Grand Peninsula G
16:30AVM-116
A method for calculating NIR bandpass-adjusted optical densities for better matching common standard test chart specifications., Christian Taylor (US); Amelia Limbocker (US) [view abstract]
Near-infrared (NIR) imaging is now prevalent in machine vision, automotive, and biomedical applications, but most step-chart definitions were created for visible imaging. Many standards assume visible lighting conditions and only consider IR-blocking, so NIR-sensitive and RGB+NIR cameras are not adequately addressed. This leads to charts whose nominal densities don't produce results as intended. We present a camera-matched methodology for designing NIR test charts whose optical densities (ODs) align with the effective bandpass of a specific camera. First, we estimate the camera illumination optics bandpass by measuring camera spectral responsivity and the measured illuminant spectrum at the sample plane. Next, we measure transmittance and/or reflectance spectra of candidate chart materials via spectrophotometry and predict their camera-effective ODs as band-integrated log-ratios. We then select step values to meet the target OD values. Validation is performed by imaging the manufactured chart in RAW, applying linearization, and comparing measured image OD to predictions and to a reference spectrophotometer integrated over the same band. The framework supports transmission and reflectance charts, mono and RGB-NIR sensors, and bands spanning ~780 1100 nm. We report a practical design recipe and guidelines for dynamic-range coverage and repeatability, enabling camera-aware NIR chart optimization rather than one-size-fits-all designs.
16:50AVM-117
Circular-edge SFR for camera image quality assessment: Theory, implementation, and validation, Xingbo Wang, Shanghai Yanding Tech. Co., Ltd (China (Mainland)); Sangkyu Yang, CIZEN TECH Co., Ltd. (Republic of Korea) [view abstract]
lang="EN-US">Spatial Frequency Response (SFR), typically measured by the ISO 12233 slanted-edge method, is a standard metric for quantifying image sharpness. Despite its robustness, the method is limited by orientation dependence, sensitivity to tilt angle, and distortion-induced edge curvature. Circular-edge SFR has been proposed as an alternative, but its properties and applicability remain largely unexplored. lang="EN-US" style="font-family: "Calibri", sans-serif; font-size: 11pt; line-height: 115%">We present an implemented circular-edge SFR algorithm and evaluate it through both simulation and physical experiments. Simulations incorporate diffraction, aberrations, sensor sampling, and distortion models, with ground-truth MTFs enabling accuracy assessment. Physical validation employs automotive camera modules. Preliminary results demonstrate strong agreement between circular-edge and slanted-edge SFR, supporting the feasibility of the approach. Further analyses examine robustness under distortion and orientation variation, with the goal of assessing circular-edge SFR as a complementary or alternative standard.
17:10AVM-118
Same scene, different pipeline: ISP impact on automotive detection at range, Tejus Vijayakumar, University of Limerick (Ireland); Ciar n Eising, University of Limerick; Brian Deegan, University of Galway; Patrick Denny, University of Limerick [view abstract]