Dissertation study 1 · Machine vision

Real-time weed perception in complex orchard rows

An attention-enhanced vision framework that detects, segments, tracks, localizes, and estimates the spatial density of intra-row weeds from a side-view camera.

Weed detections with persistent tracking identities and spatial grid overlays in an orchard row
EnvironmentApple orchards
ModelYOLOv7-Seg + CBAM
TrackingModified DeepSORT
Input1920 × 1080 video

Research challenge

Low branches, trunks, irrigation lines, shadows, and occlusion make orchard weeds difficult to see consistently.

Engineering approach

  1. 01Build and annotate a multi-species orchard weed dataset.
  2. 02Enhance YOLOv7 segmentation with a convolutional block attention module.
  3. 03Preserve weed identities across frames with DeepSORT and filtering.
  4. 04Convert masks into a 20 × 20 pixel grid for localized density estimation.

Systems in action

Weed tracking and real-time perception

The tracking sequence shows weed identities across orchard frames. The interface recording demonstrates how detections, size estimates, and nozzle assignments are brought together during field operation.

Field demonstration · 00:45Tracking weeds through a complex orchard scene

Detection masks, tracking IDs, and spatial grid overlays demonstrate real-time tracking and identity-preservation performance under branches, trunks, shadows, and occlusion.

System demonstration · 00:38Real-time perception-to-spraying interface

Live orchard imagery with weed detections, tracking IDs, spatial grids, size estimates, nozzle assignments, and spray activity logs shows how perception supports targeted spray decisions.

Why it matters

The system turns raw orchard video into stable, spatially actionable weed information for site-specific treatment.

Read the first-author journal article