Towards Real-Time PixOOD: Efficient Anomaly Segmentation for Autonomous Vehicles

2026-07-30

AI Summary

This work addresses the high computational cost of existing high-accuracy anomaly segmentation methods, which hinders their real-time deployment on embedded automotive platforms. For the first time, we accelerate the PixOOD method to real-time performance by redesigning its Neyman–Pearson criterion–based scoring mechanism and integrating TensorRT hardware-aware compilation to construct an efficient end-to-end anomaly segmentation pipeline. The proposed approach achieves substantial inference speedups while preserving accuracy, attaining 182 FPS on an NVIDIA RTX 4060 GPU and 75 FPS on a Jetson AGX Orin embedded platform—approximately 20× and 18× speedups, respectively. This advancement enables practical real-time anomaly detection for autonomous driving and railway systems.

Abstract

Real-time anomaly segmentation is essential for the safety of autonomous systems. Although recent approaches offer high accuracy, their computational cost limits their deployment on embedded hardware. This work presents an efficient and accelerated pipeline designed for both embedded and desktop platforms, targeting the autonomous driving and railway domains. The proposed approach reformulates the Neyman-Pearson scoring stage of PixOOD, a state-of-the-art out-of-distribution detection method, and deploys the full pipeline through hardware-optimized TensorRT compilation, reaching up to 182 FPS on a desktop NVIDIA RTX 4060 GPU and 75 FPS on the NVIDIA Jetson AGX Orin embedded platform, respectively 20x and 18x faster than the original baseline. The achieved results demonstrate that advanced anomaly segmentation can be efficiently deployed for onboard processing in autonomous driving and railway applications.

Problem

• real-time
• anomaly segmentation
• autonomous vehicles
• embedded hardware
• out-of-distribution detection

Innovation

• real-time anomaly segmentation
• PixOOD
• TensorRT optimization
• embedded deployment
• out-of-distribution detection