AGRICULTURAL ROBOTICS

Our custom-built Information Collection Vehicle (ICV) integrates a 16-channel 3D LiDAR, IMU, and Encoder, alongside an RTK GNSS system, to capture real-time fruit tree canopy point clouds. A rigorous Python-based workflow handles data synchronization, filtering, and registration, ensuring high-fidelity morphological trait extraction. The process includes precise scanning angle decoding, distance value validation, and kinematic state determination to discard invalid frames, culminating in reconstructed point clouds for accurate parameter extraction.

Traditional whole-tree metrics often obscure critical localized errors. This study introduces a novel 12-zone refined evaluation framework, vertically segmenting the canopy into upper, middle, and lower layers and horizontally into quadrants. This granular analysis revealed that while the 2.0 m IH consistently yielded the lowest REs in the middle layer (~2.98%), higher IHs (2.6 m) led to significant errors in the lower canopy due to blind spots and leaf phototropism, even if aggregate whole-tree REs appeared acceptable. This 'error compensation effect' underscores the need for layer-stratified assessment in precision agriculture.

Driving direction interacts significantly with LiDAR installation height, particularly at higher IHs. While low-to-medium IHs (1.4 m, 2.0 m) showed stable performance for both forward and reverse driving, the 2.6 m IH exhibited significant instability during reverse traversal (p < 0.05), especially for tree height measurements. This instability is attributed to the 'lever arm effect', where extended sensor height amplifies mechanical vibrations caused by chassis pitch/roll and altered track tension in reverse, leading to severe point cloud mismatching and degrading accuracy.