CRII: RI: A Biologically-Inspired Algorithm to Detect, Segment, and Track Moving Objects with Observer Motion

该项目旨在开发一种实时计算算法，以检测，分割和跟踪存在观察者运动的移动对象。该算法将基于脊椎动物视网膜中的对象运动敏感（OMS）细胞的实验测量特性的计算模型，其解决了类似的问题。从视网膜的已知计算特性，预计该算法将是强大的运动跟踪中的困难任务下，包括对象遮挡，多个移动对象，变化的场景统计，大量的背景运动，和光流。从生物学的角度来看，开发的计算模型将深入了解视网膜如何编码移动物体。从工程的角度来看，由此产生的计算算法将适用于移动平台的机器视觉，包括自动驾驶汽车，监视和侦察应用，以及智能传感器设计和神经形态系统。 尽管运动分析方法取得了许多进展，但与静态观测相比，基于运动观测的技术仍然处于初级阶段，就可靠性、效率、鲁棒性和运行时间而言，在真实的世界场景中。另一方面，我们的生物视觉系统能够在恒定的眼球运动的情况下可靠地执行类似的运动计算。最近，人们发现，运动物体的分割，并拒绝背景运动，开始在视网膜。视网膜神经节细胞的一个子集响应于感受野中心和周围之间的差分运动，如由在背景上移动的物体产生的，但是被全局图像运动强烈抑制，如由观察者的头部或眼睛运动产生的。这种对差分运动的选择性与物体的方向和空间模式无关，使我们的视觉系统能够找到运动物体的边界，分离多个运动物体，并预测运动的方向。视网膜使用仅五种基本细胞类型的网络同时、实时、高精度地执行这些任务。这些特性，沿着视网膜的实验可及性，使得这种神经回路成为一种理想的工作生物系统，可用作目标跟踪算法的设计。这个项目将开发一个独特的运动分析算法的移动的观察者基于视网膜的运动计算和电路的最新研究结果，并证明其在各种现实的情况下，其中包括横向观察者运动，光流，动态场景，和其他物体遮挡移动物体的性能。

This project aims to develop a real-time computational algorithm to detect, segment and track moving objects in the presence of observer motion. This algorithm will be based on a computational model of experimentally measured properties of Object Motion Sensitive (OMS) cells in the vertebrate retina, which solve a similar problem. From the known computational properties of the retina, it is expected that the algorithm will be robust under difficult tasks in motion tracking, including object occlusion, multiple moving objects, varying scene statistics, substantial background motion, and optic flow. From a biological perspective, the developed computational model will give insight to how the retina encodes moving objects. From an engineering perspective, the resultant computational algorithm will be applicable to machine vision from a moving platform, including autonomous vehicles, surveillance and reconnaissance applications, as well as, smart sensor design and neuromorphic systems. Despite many advances in motion analysis methods, techniques based on moving observations are still in a preliminary stage when compared to static observations, as far as reliability, efficiency, robustness, and runtime are concerned in real world scenarios. At other hand, our biological visual system is capable of performing similar motion computations reliably in the presence of constant eye movements. Recently, it was discovered that segmentation of moving objects, and rejection of background motion, begins in the retina. A subset of retinal ganglion cells responds to differential motion between the receptive field center and surround, as produced by an object moving over the background, but are strongly suppressed by global image motion, as produced by the observer's head or eye movements. This selectivity for differential motion is independent of direction and the spatial pattern of the object, enabling our visual system to find the boundaries of the moving objects, segregate multiple moving objects, and also anticipate the direction of motion. The retina performs these tasks simultaneously, in real-time, and with high accuracy using a network of only five basic cell types. These properties, along with the experimental accessibility of the retina, makes this neural circuit an ideal working biological system to serve as the design for an object-tracking algorithm. This project will develop a unique motion analysis algorithm for mobile observers based on recent findings of retina's motion computations and circuitry, and demonstrate its performance in a variety of realistic scenarios, which includes lateral observer motion, optic flow, dynamic scenes, and other objects that occlude the moving object.