基于LGMD建模的物体深度运动方向估计方法

An Estimation Method for the Direction of Object’s Motion in Depth Based on LGMD Modeling

摘要: 在移动机器人、自动驾驶、视频监控等应用领域，复杂的动态场景中，对于物体深度运动及方向检测一直是计算机视觉技术的难点。自然界中，昆虫在飞行过程中利用复眼视觉检测高度变化且视觉杂乱环境中的深度运动物体(或称为目标)，是学习运动感知策略的良好范例。受飞行昆虫复眼视觉功能优势的启发，本文采用物体运动的缩放变量计算与基于LGMD (lobula giant movement detector)的改进型碰撞检测模型相结合的仿生策略，提出一种基于LGMD神经元建模的物体深度运动方向估计方法(简称LGMD-ED)，通过对PPT合成动画视频和拍摄真实场景的样本视频进行仿真实验和测试，验证了本文所提新方法对于检测与估计物体远离和靠近二种典型深度运动方向的有效性。

Abstract: In the application fields including mobile robots, self-driving automobile, video surveillance, detecting the direction and motion in depth of objects in complex dynamic scenes has always been difficult in computer vision technology. In nature, the insects apply compound eye’s vision to detect a motion in depth of object (or target) in highly variable and visual cluttered environments during flight, which are excellent paradigms to learn motion perception strategies. Inspired by the advantages of the compound eye’s vision function of flying insects, this paper adopts the bionic strategy of combining the zoom variable calculation of object motion with the improved collision detection model based on LGMD (lobular giant movement detector), and proposes a method of estimating direction of object’s motion in depth based on LGMD neuron modeling (hereinafter referred to as LGMD-ED). Using the PPT composite animation videos and the sample videos of real scenes taken to do simulation experiments and tests, the effectiveness of the proposed method in detecting and estimating the objects moving away and approaching which are two typical directions of motion in depth is verified.