Rat Behavior Recognition Based On Trajectory Analysis

Moving object recognition is one of the research highlights in the field of Computer Vision. It has extensive use in a lot of fields like Security, Military, Navigation, Monitoring, Research etc. With the development of computer science and the National Internet Plus strategy, there will be more and more use for video detection and computer vision algorithms. Moving object behavior recognition is a very important branch in this field.In this paper, we take rat-robot as out research object, and focus on the problem of moving object recognition and behavior recognition based on trajectory. For object recognition, we use Vibe algorithm for background module and foreground detection. The experimental results show an excellent improvement in foreground detection compared with the plain background subtraction algorithm. For the problem of behavior recognition, we propose a method that can remove noise from trajectory data points based on Simulated Annealing algorithm. Based on that, we extract geometric curvature feature of trajectory and propose a behavior recognition algorithm base on curvature feature and probability method. Specifically, this paper includes the following aspects:1) Use the Vibe algorithm to split the background and foreground of rat video in a very efficient way without losing accuracy. This method can also overcome noise interference like illumination variation, camera vibration.2) By analyzing the foreground of rat video, we extract the skeleton and center in every frame. Then, we extract the main trajectory points by Simulated Annealing algorithm. If we need reappear the trajectory, we can use B-Spline to fit the main trajectory points. This method can save more storage space, overcome most of the noise interference, and in the same time, make data’s usability better.3) Propose an Adaptive map cover algorithm by analyzing rat’s trajectory and its sequential characteristic. Then we build a naive Bayes module to do behavior recognition and get a good score. After that, we build a classification module based on geometric curvature sequence feature. This module combined Support Vector Machine and Probability method together to improve the classifier’s performance.