FE-DETRAC

With the advance of AI, road object detection and multi-object tracking (MOT) have been prominent topics in computer vision, mostly relying on perspective cameras. Fisheye lenses provide omnidirectional wide coverage, allowing a single device to cover multiple lanes and approaches, significantly reducing hardware and installation costs. However, the inherent radial projection introduces strong view distortions, scale shrinkage near the periphery, and frequent occlusions, which severely degrade detection and, more critically, disrupt data association in tracking tasks. To our knowledge, there is a distinct lack of large-scale open benchmarks tailored specifically for infrastructure-based fisheye traffic surveillance.

This page introduces the FE-DETRAC benchmark dataset, a comprehensive extension of our preliminary FishEye8K, designed for both urban traffic detection and tracking. The released benchmark contains 20,000 continuous video frames annotated with over 470,000 bounding boxes across five road-user categories (Pedestrian, Bike, Car, Bus, and Truck). The dataset was captured by 22 fixed fisheye IoT cameras at real intersections in Hsinchu City, Taiwan, at high resolutions of 1920x1080 and 1920x1920.

The data annotation and validation processes were arduous and time-consuming due to the extreme scale variations and panoramic hemispherical distortions, particularly for small, distant objects like people riding scooters. To facilitate versatile research, the dataset is provided in multiple annotation formats, including Pascal-VOC, COCO, MOT, and YOLO. To ensure rigorous evaluation and avoid bias, we implemented camera-disjoint splits for the training and test sets.

In addition to the dataset, this work proposes a novel Hybrid Data Association (HDA) module to resolve non-linear motion mismatches without computationally expensive full-frame dewarping. We present extensive benchmark results for State-of-The-Art (SoTA) detection models (including variations of YOLOv5, YOLOR, YOLOv7, and YOLOv8) and MOT algorithms (such as StrongSORT and BoT-SORT-R). Experimental results demonstrate that integrating our HDA module significantly improves tracking performance (HOTA metrics) while retaining real-time feasibility on edge devices.

Copyright

The FE-DETRAC dataset described on this page is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 License, which implies that you must: * (1) attribute the work as specified by the original authors, * (2) may not use this work for commercial purposes (for commercial use, please contact us), and * (3) if you alter, transform, or build upon this work, you may distribute the resulting work only under the same license. The dataset is provided "as it is" and we are not responsible for any subsequence from using this dataset.