Systems and Methods for Video-Based Orchard Item Counting and Fruit Weight Estimation Using Motion-Guided Tracking and Elliptical Volume Modelling

Abstract

Abstract A computer-implemented system and method are provided for counting orchard items, including winter buds, flower buds, flowers, fruitlets, and fruit, and estimating volume, weight, and / or yield from sampled video frames selected according to at least one of time, distance travelled, location data, or motion-sensor data during row-wise orchard scanning by a moving handheld or vehicle-mounted device. Orchard items that are substantially stationary relative to orchard structure are associated across frames using a scene-level, non-object-specific camera-induced inter-frame displacement representing expected pixel-space translation of stationary scene elements and configurable distance-based gated association with gate size determined from at least frame rate, expected camera speed, orchard row geometry, or scene depth variation, while unique identifiers are retained through configurable missed sampled frames so that each item is counted once during a visible lifespan within a scanned orchard segment comprising a row, bay, or block. For fruit weight estimation, multiple frames associated with a same unique identifier and representing different views of the same fruit item are used to generate polygon masks and fitted ellipses, observable minor axes from the fitted ellipses are aggregated to derive an image-observed minor-axis characterisation, corresponding major axes are determined from population-derived characterising data obtained from laboratory measurements of representative fruit items rather than directly extracted as primary inputs from the fitted ellipses used in the method, and ellipsoidal or spheroidal volume estimates based on the image-observed minor-axis characterisation and the determined major axes are converted to weight and yield outputs. Spatial outputs such as count maps, density maps, weight maps, and yield forecasts may be generated. 20 26 20 31 24 27 A pr 2 02 6 2 0 2 6 2 0 3 1 2 4 2 7 2 0 2 6 A p r 2 0 2 6 2 0 3 1 2 4 2 7 2 0 2 6 A p r

Claims

1. A computer-implemented method of counting orchard items in an orchard from a sequence of video frames captured during row-wise scanning of the orchard by a handheld or vehicle-mounted camera moving relative to substantially stationary orchard items, the method comprising:detecting, in sampled video frames selected according to at least one of time, distance travelled, location data, or motion-sensor data, orchard items in a first frame and in a subsequent frame using a neural-network detector, and generating, for each detected orchard item, spatial location data comprising a centroid reference point derived from a bounding box or segmentation mask;determining, between the first frame and the subsequent frame, a scene-level, non-object-specific camera-induced inter-frame displacement representing expected pixel-space translation of stationary scene elements arising from movement of the camera relative to the substantially stationary orchard items, wherein the scene-level camera-induced inter-frame displacement is determined using optical flow, feature displacement analysis, frame alignment, block matching, or a combination thereof;for each detected orchard item in the first frame, generating a predicted expected position in the subsequent frame by translating the centroid reference point using the scene-level camera-induced inter-frame displacement;defining, in image coordinates of the subsequent frame, a configurable distance-based gated region around the predicted expected position, wherein a size of the gated region is determined from at least frame rate, expected camera speed, orchard row geometry, or scene depth variation for the row-wise scan;associating a detected orchard item in the subsequent frame with the detected orchard item in the first frame when the centroid reference point of the detected orchard item in the subsequent frame satisfies a proximity criterion relative to the gated region, wherein the association is determined primarily by spatial consistency of the scene-level predicted inter-frame displacement rather than by visual appearance similarity, wherein appearance embeddings, re-identification features, or feature similarity scores, if used, are used only as secondary or tie-breaking inputs, and wherein, when multiple detected orchard items satisfy the proximity criterion, association is resolved by nearest distance or a secondary spatial heuristic;maintaining an existing unique identifier for an associated orchard item, creating a new unique identifier when no existing unique identifier satisfies the proximity criterion, retaining an existing unique identifier through a configurable number of successive sampled frames when no associated orchard item is found, and terminating the existing unique identifier when no associated orchard item is found beyond the configurable number of successive sampled frames; andgenerating, for a row, bay, or block, an orchard-item count by counting each unique identifier once during a visible lifespan of the corresponding orchard item within the scanned orchard spatial segment.

2. The method of claim 1, wherein the orchard items comprise winter buds, flower buds, flowers, fruitlets, fruit, or any combination thereof.

3. The method of claim 1, wherein the moving device is a handheld device, a smartphone, a vehicle-mounted camera, or another field-deployable capture device.

4. The method of claim 1, wherein the gated region is circular, elliptical, rectangular, or polygonal.

5. The method of claim 1, wherein the scene-level camera-induced inter-frame displacement is a global inter-frame displacement representing translation of stationary scene elements and is not object-specific.

6. The method of claim 1, wherein the sampled video frames are selected according to distance travelled, location data, motion-sensor data, or a combination thereof.

7. The method of claim 1, further comprising receiving location data captured during the row-wise scanning, correcting or refining the location data, and fusing the corrected or refined location data with the unique identifiers to generate a spatial dataset indicating orchard-item position or orchard-item density across rows, bays, or blocks.

8. A computer-implemented system for generating spatial orchard count outputs, the system comprising one or more processors and memory storing instructions which, when executed, cause the system to:receive video frames and location data captured during row-wise scanning of an orchard by a handheld or vehicle-mounted device moving relative to substantially stationary orchard items;detect orchard items in sampled video frames selected according to at least one of time, distance travelled, location data, or motion-sensor data, determine a scene-level, non-object-specific camera-induced inter-frame displacement between successive frames representing expected pixel-space translation of stationary scene elements using optical flow, feature2026203124   25 Jun 2026displacement analysis, frame alignment, block matching, or a combination thereof, generate predicted expected positions of previously detected orchard items in subsequent frames using the scene-level camera-induced inter-frame displacement, define configurable distance-based gated regions around the predicted expected positions, wherein a size of each gated region is determined from at least frame rate, expected camera speed, orchard row geometry, or scene depth variation, and assign and maintain unique identifiers for the orchard items by proximity-based association determined primarily by predicted spatial displacement rather than by visual appearance similarity, including retention of a unique identifier through a configurable number of successive sampled frames lacking an association, wherein appearance embeddings, re-identification features, or feature similarity scores, if used, are secondary only, and wherein, when multiple detected orchard items fall within a respective gated region, association is resolved by nearest distance or a secondary spatial heuristic;correct or refine the location data to reduce drift, jitter, or orchard-specific distortion and fuse the corrected or refined location data with the unique identifiers to generate a spatial dataset indicating orchard-item position and orchard-item density across rows, bays, or blocks;generate output data representing orchard-item count and orchard-item density for a row, bay, block, or other defined unit; andpresent the output data in a dashboard or other user interface for a grower or orchard manager.

9. A non-transitory computer-readable medium storing instructions which, when executed by one or more processors, cause performance of the method of any one of claims 1 to 7.

Images