Depth-Camera Point-Cloud-Driven MambaMixer for Orchard Branch Obstacle Perception

Abstract

 Before the manipulator, sprayer boom or mobile platform enters the unsafe clearance area, an autonomous orchard vehicle needs to be able to detect a thin branch. Depth cameras are relatively low-cost metric geometry sources; however, leaf occlusion, mixed pixels, range noise, and irregular sampling can break branch continuity. This paper introduces a point cloud-based MambaMixer, which combines geometry-aware tokenization with interleaved spatial and scale order selective scanning. Local covariance descriptors retain cylindrical evidence, bidirectional Morton traversal guarantees continuity of separated observations, and a confidence gate integrates global context with fine-scale diameter cues. Semantic, centerline offset, radius, and gap head are jointly trained under topological-aware supervision. A new orchard dataset has 18,420 synchronized depth frames, 2.31 billion valid points, four tree architectures, three depth cameras, and branch diameters from 6 to 94 mm. The model has reached 91.8% branch intersection-over-union, 94.6% obstacle recall, 18.7 mm centerline error, and 12.9 mm clearance error. It processes a 65,536-point frame in 28.4ms with 3.7GB of memory, exceeds the best point-transformer baseline by 3.5% points in intersection-over-union, and reduces latency by 37.0%. 45% structured point loss; recall is still 88.1%. Based on the above results, linear-complexity state-space mixing can maintain thin-branch geometry and offer clearance estimates suitable for real-time orchard motion safety.