Browse papers

3 papers

Evidence-Driven Multi-Sensor Monitoring for Railway Track Safety

Danush M · Aug 15, 2026

Abstract : Railway safety systems operate in an environment where a missed obstacle can have severe consequences, while excessive false alarms can reduce the usefulness of an alerting system. Vision systems provide rich semantic information but can degrade under poor lighting or adverse weather. Radar offers complementary distance and motion information and is less dependent on visible illumination, while other short-range sensors can provide inexpensive confirmation signals. This paper proposes an evidence-driven framework, called Virtual Dome, for combining heterogeneous sensors into a layered railway track-monitoring system. The central idea is not to treat any single sensor or machine-learning model as an authority. Instead, the framework separates sensing, feature extraction, evidence fusion, confidence estimation, and alert generation. A track event is escalated when multiple independent signals support the same interpretation or when a high-risk signal persists beyond a defined temporal window. The paper also proposes a practical evaluation protocol covering detection performance, false alarms, latency, environmental robustness, sensor disagreement, and computational cost. Particular attention is given to data scarcity and the difficulty of representing mixed human, animal, object, and infrastructure events in a single dataset. The resulting framework is intended as a research and prototyping architecture rather than a claim of deployment readiness.

Cyber-Physical SystemsSensor FusionComputer VisionRailway SafetyArtificial Intelligence

Technical Feasibility Analysis of a Low-Cost UAV-Based Computer Vision System for Wildlife Monitoring

jeevanr17 · Aug 15, 2026

Abstract. Wildlife monitoring in forested and environmentally challenging regions requires observation systems that can provide sufficient coverage while reducing the dependence on continuous manual surveillance. Unmanned aerial vehicles (UAVs) combined with computer vision provide a potential approach for automated wildlife observation; however, practical deployment is constrained by factors such as payload capacity, propulsion requirements, wireless communication, computational resources, environmental conditions, and detection performance. This paper presents a technical feasibility analysis of a low-cost UAV-based wildlife monitoring system integrating a Pixhawk 2.4.8 flight controller, GPS-assisted navigation, an ESP32-CAM for image acquisition, Wi-Fi-based video transmission, and YOLOv8s for animal detection. The detection model was trained for ten wildlife classes using a dataset containing 5,553 images and 8,014 annotated object instances, with 4,442 images used for training and 1,111 for validation. The model was trained for 100 epochs at an image size of 640 × 640 pixels with a batch size of 8. At the final training epoch, the model achieved a validation precision of 94.01%, recall of 94.01%, mAP@50 of 96.76%, and mAP@50–95 of 83.54%. The system was also examined at the system level with respect to wireless communication, image acquisition, and environmental conditions. The UAV configuration uses A2212/13T 1000-KV brushless motors, 10 × 4.5 propellers, and an 11.1-V 5400-mAh Li-Po battery, with a UAV mass of approximately 2.0 kg and a considered payload range of 1.2–1.5 kg. Engineering calculations are used to examine the propulsion requirements for the resulting take-off mass, while flight endurance is treated as a theoretical parameter because an experimentally measured endurance value was not established. The analysis evaluates the technical feasibility of integrating low-cost UAV hardware with external computer-vision processing while identifying communication dependence, low-light degradation, and the absence of onboard AI processing as important limitations.

YOLOv8RoboticsUAVComputer VisionWildlife Monitoring

Skin Cancer Detection Using Preprocessed Vision Transformers: An Interpretability-Driven Approach

Ananya Atri · Aug 15, 2026

Abstract : Skin cancer is an issue affecting health all over the world, and proper diagnosis of the issue would lead to better patient outcomes. Deep Learning, in particular Convolutional Neural Networks (CNNs), has brought more objective and fast screening techniques. This paper assesses the diagnostic capability of Vision Transformers (ViT) with Explainable AI (XAI) on HAM10000 data (10,015 pigmented skin lesion images). The researchers used a critically important preprocessing pipeline comprising of morphology transformations to remove hair, image segmentation to identify regions of interest and data augmentation to overcome imbalanced classes. Five XAI techniques were used, namely, Grad-CAM, LIME, SHAP, Integrated Gradients and Saliency Maps to improve diagnostic transparency. The experiment outcomes revealed that ViT + Grad-CAM architecture was the best performer with a total accuracy of 96.6%. The model achieved 0.92 precision, 0.90 recall and 0.92 F1-score. Grad-CAM was especially useful in a clinical setting as it is fast to run and localizes classes accurately.

Skin CancerMedicineComputer VisionArtificial IntelligenceVision Transformer