Shanghai University of Engineering Science China.
Shanghai University of Engineering Science China.
Behav Processes. 2024 Oct;222:105109. doi: 10.1016/j.beproc.2024.105109. Epub 2024 Sep 25.
Collective animal behavior occurs in groups and swarms at almost every biological scale, from single-celled organisms to the largest animals on Earth. The intriguing mysteries behind these group behaviors have attracted many scholars, and while it is known that models can reproduce qualitative features of such complex behaviors, this requires data from real animals to demonstrate, and obtaining data on the exact features of these groups is tricky. In this paper, we propose the Hidden Markov Unscented Tracker (HMUT), which combines the state prediction capability of HMM and the high-precision nonlinear processing capability of UKF. This prediction-driven tracking mechanism enables HMUT to quickly adjust tracking strategies when facing sudden changes in target motion direction or rapid changes in speed, reducing the risk of tracking loss. Videos of fruit fly swarm movement in an enclosed environment are captured using stereo cameras. For the captured fruit fly images, the thresholded AKAZE algorithm is first used to detect the positions of individual fruit flies in the images, and the motion of the fruit flies is modeled using a multidimensional hidden Markov model (HMM). Tracking is then performed using the Unscented Kalman Filter algorithm to obtain the flight trajectories of the fruit flies in two camera views. Finally, 3D reconstruction of the trajectories in both views is achieved through polar coordinate constraints, resulting in 3D motion data of the fruit flies. Additionally, the efficiency and accuracy of the proposed algorithm are evaluated by simulating fruit fly swarm movement using the Boids algorithm. Finally, based on the tracked fruit fly flight data, behavioral characteristics of the fruit flies are analyzed from two perspectives. The first is a statistical analysis of the differences between the two behaviors. The second dimension involves clustering trajectory similarity using the DTW method based on fruit fly flight trajectories, further analyzing the similarity within clusters and differences between clusters.
群体动物行为几乎在每个生物尺度上都会发生，从单细胞生物到地球上最大的动物。这些群体行为背后的迷人奥秘吸引了许多学者，虽然人们知道模型可以再现这些复杂行为的定性特征，但这需要来自真实动物的数据来证明，并且获取这些群体的确切特征的数据很棘手。在本文中，我们提出了隐马尔可夫无迹跟踪器（HMUT），它结合了 HMM 的状态预测能力和 UKF 的高精度非线性处理能力。这种预测驱动的跟踪机制使 HMUT 能够在面对目标运动方向的突然变化或速度的快速变化时快速调整跟踪策略，从而降低跟踪丢失的风险。使用立体摄像机拍摄封闭环境中果蝇群运动的视频。对于捕获的果蝇图像，首先使用阈值 AKAZE 算法检测图像中单个果蝇的位置，并使用多维隐马尔可夫模型（HMM）对果蝇的运动进行建模。然后使用无迹卡尔曼滤波算法进行跟踪，以获得两个相机视图中果蝇的飞行轨迹。最后，通过极坐标约束实现两个视图中轨迹的 3D 重建，从而获得果蝇的 3D 运动数据。此外，通过模拟使用 Boids 算法的果蝇群运动来评估所提出算法的效率和准确性。最后，基于跟踪的果蝇飞行数据，从两个角度分析果蝇的行为特征。第一个角度是对两种行为之间差异的统计分析。第二个维度涉及基于果蝇飞行轨迹使用 DTW 方法对轨迹相似性进行聚类，进一步分析聚类内的相似性和聚类之间的差异。
