Multi-object extraction from topview group-housed pig images based on adaptive partitioning and multilevel thresholding segmentation

• Multi-object extraction method for overhead views of group-housed pigs is proposed.
• Adaptive partitioning and multilevel thresholding segmentation is used.
• Circular sub-block radiuses are obtained using centroid-edge point distance curves.
• The method achieves an average detection rate of 92.5% and operates in real time.

The aim of this study is to provide a feasible method that can accurately extract individual pigs from a drinker and feeder zone; therefore, an object extraction method based on adaptive partitioning and multilevel thresholding segmentation is proposed. First, a single frame image is enhanced using histogram equalisation, and then it is segmented with a maximum entropy global threshold. The initial segmentation objects are obtained by extracting a “valid area” and morphological processing. Then, each object centroid is calculated from the initially segmented objects, and the original image is adaptively divided into multiple circular sub-blocks whose origin is the centroid and radius is the maximum distance from the centroid to the edge point. Finally, an accurate secondary segmentation result is obtained using multilevel thresholding segmentation in each sub-block. The test data included thirty random videos collected in AVI format, and 9000 frames from 5 days × 6 videos × 120 s × 25 frames s−1 were selected. Results show that the average detection rate is 92.5%. This paper also analyses the possible applications of the proposed method to pig behaviour analysis, individual recognition, and weight estimation.