Tracking Zebrafish in 3D using Stereo Vision
Authors
Pedersen, Malte ; Bengtson, Stefan Hein
Term
4. term
Publication year
2017
Submitted on
2017-06-08
Pages
100
Abstract
Sebrafisk (Danio rerio) bruges i stigende grad i genetik-, miljø- og lægemiddelforskning. For at forstå adfærd analyserer man ofte fiskenes bevægelser, hvilket kræver computersyn, der kan spore dem automatisk. De fleste tilgængelige værktøjer er dog udviklet til landdyr som mus og rotter og sporer typisk kun én fisk eller kun i to dimensioner. Denne afhandling undersøger mulighederne for at spore sebrafisk i tre dimensioner. Det foreslåede system anvender hyldevarer: to GoPro HERO5 kameraer i et stereosyn-setup. I hver video opdages fisk med SURF-nøglepunktsdetektoren, en metode der finder karakteristiske punkter i billedet. De tredimensionelle positioner estimeres ved at kombinere de to synsvinkler. Fordi lys brydes, når det passerer mellem luft og vand, baseres 3D-rekonstruktionen på strålesporing sammen med Snells lov for at korrigere for brydning ved grænsefladen. For at følge fisk over tid kobler systemet observationer på tværs af frames med et Kalman-filter, som forudsiger jævn bevægelse, og Munkres-algoritmen (også kaldet den ungarske), som matcher observationer til spor. Derved samles korte forløb (tracklets) med en gennemsnitlig længde på ca. 150 frames. En afsluttende komponent, der skulle kæde disse delspor sammen til fulde forløb, er ikke implementeret i denne version.
Zebrafish are widely used in genetics, environmental, and drug research. To study behavior, researchers often analyze how the fish move, which requires computer vision systems that can track their positions automatically. Most available tools, however, were built for land animals like mice and rats and usually track only one animal or only in two dimensions. This thesis explores how to track zebrafish in three dimensions. The proposed setup uses off-the-shelf hardware: two GoPro HERO5 cameras arranged as a stereo vision pair. In each video, fish are detected with the SURF keypoint extractor, a method that finds distinctive points in the image. Their 3D positions are then estimated by combining the two views. Because light bends when it passes between air and water, the 3D reconstruction uses ray tracing together with Snell's law to correct for refraction at the interface. To follow fish over time, the system links detections across frames using a Kalman filter, which predicts smooth motion, and the Munkres (Hungarian) algorithm, which matches detections to tracks. This assembles short trajectory segments (tracklets) with an average length of about 150 frames. A final component that would link these tracklets into complete trajectories was not implemented in this iteration.
[This abstract was generated with the help of AI]
Documents