FOCRD: Fully Occluded Citrus Radar Dataset

In this dataset, a comprehensive collection of high-fidelity radar signatures is presented to address the critical challenges of autonomous citrus perception. This dataset was generated during a 45-day intensive data collection period at the University of Wollongong, within the Applied Mechatronics and Biomedical Engineering Research (AMBER) group. All scans were performed in a controlled laboratory environment using real citrus fruits and genuine foliage to capture precise radar responses for agricultural robotics research. The primary goal of this collection is to provide a reliable sensing alternative for citrus detection and localisation where traditional vision-based systems fail due to 100% visual occlusion. The dataset architecture is fully compatible with shuffling and curriculum learning strategies to support diverse model training requirements.

Dataset Organisation (Curriculum Phases):The data is organised into three complexity levels to support hierarchical learning strategies as defined in the dataset's splitting configuration:

Phase 1 (Easy): Consists of unoccluded citrus and background samples.

Phase 2 (Medium): Includes unoccluded citrus along with only_leaf and leaf+branch negatives.

Phase 3 (Hard): Represents the full set and introduces the most challenging scenarios with fully occluded citrus targets hidden behind dense foliage, alongside background, only_leaf, and leaf+branch samples.

Technical Acquisition & Format:

Sensor & Hardware: 60GHz Acconeer A121 pulsed coherent radar mounted on a custom 2-DOF pan-tilt system, covering a 50° azimuth and 35° elevation field of view.

Data Processing: Raw IQ signals are processed into 36 × 51 radar intensity images using a temporal-variance filter to suppress dynamic reflections from foliage.

Metadata Sync: Each of the 931 unique scans is synchronized with a 4D metadata vector: [Grid ID, Total Distance, Sensor-to-Leaf Distance, Foliage Freshness].

Regional Compensation: Grid ID covers 9 sectors (16.6° × 11.6°) to account for angular leakage.

Environmental Parameters: Includes an angular step size of 1°, range resolution of 0.0025 m across 160 bins, and leaf thicknesses ranging from 0.8 to 1.4 mm.

Data Partitioning & Augmentation:

Splits: Training, validation, and testing subsets are partitioned in a 50:25:25 ratio.

Augmentation: Training data is augmented through offline horizontal flipping, involving a synchronized transformation of intensity maps, bounding-box coordinates, and spatial Grid IDs to maintain physical context consistency across the symmetric azimuth plane.

Significance:The inclusion of the 4D context vector allows for a deep analysis of environmental disturbers and signal attenuation, offering a solution to the fundamental challenges of heavy occlusion in orchard environments. This structure makes the dataset suitable for testing robotic perception in real-world agricultural tasks such as precise harvesting and targeted spraying.