Automatic Tracking System Of Zebra Fish School

Animal aggregation exists prevalently in nature and this fascinating phenomenon has long attracted a great deal of attention from scientists in various fields striving to discover the underlying rules. What complex behavior these seemingly simple individuals can generate! What’s the hidden mechanism? What’s the role of these behaviors and how do they work? By looking for the answers to these questions, not only can it help us explore and understand our nature but also apply the discoveries to human society to promote its development. Zebra fish is a model organism and widely used as a subject for studying such kind of behavior. It has become the main method of studying collective behavior by obtaining the motion trajectories of zebra fish school and analyzing each trajectory and the relationship among trajectories. Video camera provides a convenient choice to get their trajectories.Through recently camera system and technique of computer vision have developed greatly, getting the trajectory of zebra fish precisely and robustly from video is still very challenging due to:(1)zebra fish’s body has a volume and is very flexible, which can be simply treated as a point; (2)there are severe and frequent occlusion or interaction among fish in the case of tracking a large population fish.Motivated to solve these challenges, the thesis proposes two automatic methods of tracking zebra fish school. Each method formulates the problem from different views and has advantages respectively.The first fish tracking method is based on detection. According to the prominent characteristic of fish’s head and its imaging feature, we design a learning-based head detector. This detector not only can get the accurate position and orientation of head but also can avid the effect of body’s deformation. We propose a tracking method that models the motion of zebra fish and combines a global assignment algorithm, which makes the tracking method robust against missing detection and false detection.The second one is contour tracking method of fish school. We model the fish body by a subspace representation which is learned from hundreds of manually collected samples. which not only can precisely describe fish deformation but also has fewer parameters. Tracking is thus greatly facilitated as much fewer parameters need to be estimated. Instead of estimating the parameters of all fish simultaneously which may result in an extremely high dimensional state space, we introduce the notion "group" by dividing all targets into groups according to their mutual distance. Tracking in each group is independent and parallel. Parameters of targets in the same group are estimated jointly with a novel effective cost energy to model their occlusion. The complex change of groups’composition is dealt with by a simple decomposed two-steps method.The two methods proposed by this thesis can solve the challenges of zebra fish school tracking and have advantages respectively. The first method is simple and high speed and the second one can provide more information for trajectory analysis such as the motion state of body. Excessive experimental results demonstrate the accuracy and robustness of the two proposed methods.