Unsupervised Learning of Motion Patterns for Object Classification in Aquaculture
MetadataVis full innførsel
In recent years, computer vision has been used in increasing amounts in aqua culture and will be essential for the development of automated solutions for fish farming. In this thesis we study the possibility of using unsupervised learning based on motion patterns to classify main groups of objects in a fish farm. The main focus will be separating fish from feed. The approach is based on the hypothesis that fish and feed have distinct motion patterns that are suitable to use as classification criteria. The approach is based on optical flow by using KLT-tracking to estimate motion in the image. Similar motion patterns are grouped together using cluster analysis. Mean shift and DBSCAN were chosen as the algorithms to be used in the experiments, based on a preliminary analysis of the motion data. Mean shift is centroid based, while DBSCAN is density based which gives a useful combination of differing properties to compare. Further, the effect of adding object sizes to the clustering was studied. Object sizes were estimated by using image segmentation. The segmentation algorithm is based on edge detection, using a Sobel operator to create gradient images that can be used as basis for finding contours of objects. Results showed that classifying fish and feed based on motion patterns is plausible under certain conditions. There are some requirements for the camera position that improves the classification. For instance the clustering performance increases when numerous objects is visible in the image. Accurately selecting clustering parameters are also necessary to avoid cluster merging. In cases where several clusters are merged together, all valuable information about the objects are lost. The effect of adding object sizes to the clustering proved to be as expected. It resulted in improved separability of the motion patterns, although the segmentation accuracy required, made the proposed approach not robust enough to calculate the object sizes automatically. The main issues were incomplete contours because of too low contrast towards the background and overlapping objects. As a cause of the inaccurate segmentation, the number of data samples were too small to draw any conclusions, although the tendencies are that by using a better suited segmentation algorithm, a more consistent classification can be obtained.