Detecting cattle: thermal-RGB fusion
Counting and monitoring cattle from a drone is a fusion problem: thermal imagery makes warm bodies stand out but suffers from similar backgrounds and occlusion, while RGB supplies the colors and textures that thermal lacks. The group’s fusion-based detection framework passes preprocessing, feature extraction and feature fusion into multiple object detectors, then applies a new non-maximum suppression step. Tested on drone video, the deep-learning model correctly reports “10 cows detected.”
Thermal and RGB aerial views of the same pasture.
The work is grounded in the field with partner LandView (Markus Weber, President). The company was born out of the work and staff of Serecon Inc., the Alberta-based consulting firm focused on farmland valuation and agricultural management consulting.
Predicting weight from a single image
The same imagery also answers a rancher’s next question: how much does each animal weigh? From a single-view image, the pipeline segments the animal, infills occluded parts, extracts the torso and body, and records body measurements; geometrical and mass feature descriptors then feed a Three-Stream DCNN that predicts the weight: 294 kg in the example output.
Segmentation, infill and feature extraction feed a Three-Stream DCNN: 294 kg predicted.