Back to the scenes
Pek Automotive, Apr 2025 to present
Depth and detection for clustered fruit
Replaced peak-based depth with a front-surface search, made exposure follow changing sunlight, and retrained YOLO segmentation on occluded clusters from 700 GB of field recordings.
700 GB of recordings
Problem
The robot measures fruit depth with an Intel RealSense D435. Taking the peak of the depth values inside a fruit’s mask overestimated depth systematically, because of the fruit’s round shape and the foliage around it. Outdoors, changing sunlight overexposed the infrared and depth frames. And the detector struggled with fruit partly hidden in a cluster.
Approach
- Depth. A histogram front-surface search replaces the peak, and fruit pixel masking keeps foliage out of the estimate. I also corrected the perspective-projection bias in fruit diameter estimation.
- Exposure. A saturation-based controller adapts faster than the default auto-exposure when sunlight changes.
- Detection. A semi-automated annotation workflow over 700 GB of field recordings, then YOLO instance segmentation models retrained on occluded fruit clusters.
Result
Less systematic overestimation of fruit depth, fewer overexposed infrared and depth frames, and better detection of partially visible fruit.