Session 4B – Computer Vision
Date: Thursday, 19th November 2026
Time: 09:45 – 10:45
Room: Aula Capella
Session Chair: TBA
Presentation format: Each paper is allocated 15 minutes: approximately 10–12 minutes for the presentation and 3–5 minutes for questions.
09:45 – Unsupervised Video Anomaly Detection: A Spatio-Temporal Approach using Recurrent Autoencoders
Authors: Claudio Mannarano¹, Nerea Gallego², Ana Cristina Murillo² and Eduardo Montijano²
Affiliations: ¹ Computer Science Department, Università degli Studi di Torino, Torino, Italy; ² DIIS, I3A, Universidad de Zaragoza, Zaragoza, Spain.
Abstract
The ability to automatically detect errors and anomalous events in robot executions is a fundamental requirement in safety-critical applications. Video anomaly detection (VAD) provides a vision-based approach to detecting anomalous events by modeling normal visual behavior and identifying deviations from it. However, transferring VAD techniques to robotic applications is still an open problem. On the one hand, the scarcity and diversity of failure examples make supervised training difficult, motivating methods that learn exclusively from normal executions. On the other hand, robotic anomalies frequently correspond to high-level execution errors rather than visual abnormalities. This paper addresses both challenges by proposing a new fully unsupervised VAD algorithm designed to identify time-dependent mistakes. Our algorithm contains three key elements: a video pre-processing stage to extract task-relevant visual information, a recurrent autoencoder architecture to learn normal task dynamics and a training and evaluation strategy that automatically determines the optimal anomaly detection threshold. Evaluation is performed using two distinct real-world scenarios, one involving physical disturbances and another containing time dependent errors in robot actions. The results show promising performance, supporting the feasibility of unsupervised video-based error detection for robotic systems.
10:00 – FastBEVFusion: Simple Multi-Modal Fusion for 3D Object Detection at 50 FPS
Authors: Tom Sanitz, Emre Gülsoylu, Simone Frintrop and Christian Wilms
Affiliations: University of Hamburg, Hamburg, Germany.
Abstract
Reliable real-time 3D object detection based on LiDAR and camera data is critical for safety in autonomous driving, where systems must accurately perceive their surroundings with minimal latency. Many state-of-the-art systems achieve high-quality detection results but fail to meet real-time constraints, limiting their practical use. To address this challenge, we propose FastBEVFusion, a simple, fast, and robust fusion system for 3D object detection. Based on modality-specific features, FastBEVFusion projects these features into semi-aligned modality-specific BEV spaces, without relying on a computationally intensive and potentially error-prone depth estimation. Given these BEV spaces, we propose a light-weight, soft, and dynamic fusion module based on cross-attention, which adapts dynamically to the scene context and enables feature fusion across the entire BEV spaces. Experiments on the complex nuScenes dataset show that FastBEVFusion outperforms state-of-the-art real-time systems, and achieves over 50 FPS without deployment libraries. Additionally, our system shows strong robustness to calibration errors, making FastBEVFusion a powerful 3D object detection system for real-world applications.
10:15 – DTF-Track – Dual-Tier Fusion for People Tracking
Authors: Edison Bejarano¹,², Valerio Bo¹,², Anaís Garrell¹,² and Alberto Sanfeliu¹,²
Affiliations: ¹ Institut de Robòtica i Informàtica Industrial (IRI, CSIC–UPC), Barcelona, Spain; ² Universitat Politècnica de Catalunya – BarcelonaTech (UPC), Barcelona, Spain.
Abstract
Mobile service robots operating among humans need reliable 2D LiDAR-based people detection and tracking when 3D sensing is unavailable, in the presence of ego-motion, clutter, and brief occlusions. SPENCER established a widely used ROS 1 pipeline that combines laser classifiers, multimodal detection, and a nearest-neighbor tracker, but its distributed nodelet graph is difficult to maintain on modern ROS 2 platforms. We present DTF-Track, a unified ROS 2 node that preserves SPENCER’s laser feature models while replacing its union-based detector fusion with consensus-based multi-detector validation, introducing dual-tier spawn/maintenance policies, LiDAR-first Kalman tracking with decoupled YOLO semantic confirmation, and temporally coherent scan-to-map projection. We test two hypotheses: that multi-detector consensus at track initiation reduces false positives relative to SPENCER’s single-detector design, and that an occlusion-aware tracking lifecycle mitigates multi-person identity fragmentation. Four scenarios using TIAGO Pro (DTF-Track) and IVO (SPENCER) support both claims in a cross-platform comparison, yielding 43–47% fewer excess IDs in clutter and 62–72% fewer during occluded crossings, with roughly twice the precision of the baseline and no significant loss of recall.
10:30 – Open-Vocabulary 3D Scene Graph construction through Object Tracking and Spatial Localization
Authors: Sara Carles-Pamias and Alberto Sanfeliu
Affiliations: Institut de Robòtica i Informàtica Industrial (IRI, CSIC–UPC), Universitat Politècnica de Catalunya (UPC), Spain.
Abstract
In this article, a pipeline for open-vocabulary 3D scene graph construction is presented. Existing scene graph generation pipelines, such as Conceptgraphs, rely on computationally intensive reconstruction and older foundational models, which result in a reconstruction-dependent pipeline and limit their applicability in real-world robotic environments. The proposed approach works with RGB-D image sequences to build a scene graph containing the objects (nodes) detected in the scene as well as their spatial relationships (edges). The framework integrates tracking into object detection and open-vocabulary models. It also relies on geometrical methods to reduce computational load and enhance determinism and removes the dependence and need for a full reconstruction by computing the object centroids as their 3D localizations. The proposed pipeline is evaluated on synthetic and real-world data for complete testing of its robustness. The results demonstrate the feasibility of combining open-vocabulary models with temporal tracking and geometric reasoning for scene graph construction. Moreover, they show that in real-world environments the method improves the results obtained with state of the art methods, which often rely solely on synthetic data for their results evaluation.