众所周知，蜻蜓在捕捉猎物时，能够侦测和追踪猎物，预判其行动轨迹，从而先发制人，防止其逃脱。而最新的研究显示，蜻蜓的这种能力或许会帮助自动驾驶汽车的安全性更上一个台阶。

澳大利亚阿德莱德大学医学院的研究中心总监Steven Wiederman博士认为，蜻蜓脑部有一类专门负责侦测目标的神经元，用以在捕猎过程中先发制人，采取行动。在研究者看来，这为提升自动驾驶汽车的视觉系统带来了新的思路。蜻蜓在捕猎时聚焦视线的能力十分强大，即便猎物潜伏在纷繁隐蔽的环境中，也能将其牢牢锁定。

这一神经元有望被应用在自动驾驶系统上，以进一步提升汽车的安全性。为此，Wiederman博士的团队已和来自瑞典隆德大学的研究小组开展了合作项目。而为了更好地展示这一技术的潜力，有一所大学的研究小组正试图运用全自动的轮式机器人平台来测试这一模拟蜻蜓探测与预判能力的感应技术，而这一测试也是上述项目中的重要环节。

澳大利亚的研究人员发现，蜻蜓脑中用于目标侦测的神经元，能够提升蜻蜓对所追赶猎物前方焦点区域的反应能力。这样，即使猎物忽然从蜻蜓的视野中消失，焦点也会随着时间推移而进行相应的变化，使得蜻蜓的大脑能够预测出猎物可能的运动轨迹以及接下来出现的方位，从而重新定位目标。

目前，自动驾驶汽车所面临的一大难题，就是在给定路况的情况下（对蜻蜓而言即为“给定猎物”的情况），从至少两个选项中做出优先目标选择的决策。在SAE《国际汽车工程》杂志的专访中，Wiederman博士向我们解释了昆虫学对此项研究带来的帮助。

 “在专注追踪目标，不受周边情况干扰方面，生物的大脑往往具有很强的能力。我们在蜻蜓的大脑中就发现了这样一类神经元，具有“选择性关注”的生物特征。当它同时面对两个移动目标时，神经元会选择只关注其中一个，有时甚至还能在中途切换关注目标。”

“有时神经元会下达指令，锁定一个并不那么起眼的目标。”Wiederman博士特别指出，“我们目前正在做的，就是要弄清究竟是哪种属性决定了蜻蜓追踪目标的选择？是和时机有关，还是因为目标物运动轨迹的缘故，抑或是因为该目标更值得关注？这一选择是否只和目标物的特性相关？还是说蜻蜓每次目标锁定，都受其大脑内部某种高阶活动的支配？另外，什么时候是适合锁定目标的时机？什么时候又该更换到另一更值得关注的目标？这些问题都有待我们去回答。”

通过研究这一相对可控的模型系统，研究人员希望能进一步洞悉更复杂的大脑结构是如何锁定目标的，比如：人脑如何在驾车时，如何从诸多目标物中选择并进行锁定。因此，这些研究人员正在基于蜻蜓的选择性处理模式，积极研发针对可自助移动平台的相关模型。

独特的机器人测试平台

Wiederman博士除在高校任职外，还供职于澳大利亚研究委员会（ARC）科研中心，主攻纳米级生物光子学，是视觉生理学与仿生机器人实验室（the Visual Physiology and Neurobotics Laboratory）的负责人。据他称，之所以选择蜻蜓进行研究原因有三。首先，蜻蜓是电生理学进行数据记录研究时的理想实验对象；其次，蜻蜓是自然界最高效的捕猎者之一；第三，蜻蜓所展现出的“高阶”处理能力，令人十分好奇，比如其预判和选择性专注的能力，就是其他很多低阶昆虫（如家蝇）所不具备的。

那么，蜻蜓的眼睛和人类相比，究竟有无相似之处呢？Wiederman博士解释道，蜻蜓的复眼有数千个独立的晶体，可以共同把光聚焦到某一片视网膜上。相比之下，人眼则只有一个晶体可用于相同的聚焦活动。蜻蜓复眼的分辨率很低（单眼视觉灵敏度仅为约0.8o），而人眼因为有视网膜中央凹，所以灵敏度非常高。蜻蜓双眼视域的灵敏度也仅为10o左右，且双眼的间距太近，因此要产生纵深感，就必须采取其他方式（如动态视差）来追踪目标。

然而，Wiederman博士也指出，如果比较神经元处理视觉信息时的基本方式，放眼整个动物界，不同物种之间还是有很多相似之处的。而通过运用人类精神物理学的实验方式，研究小组可以弄清何种类型的处理方式在人类和蜻蜓身上都能起到重要作用。

相关研究是在Clearpath Robotics公司的Husky A200型无人地面车上进行的。这一配备了摄像头的机器操作平台采用的是开源串行协议。经大学研究小组的设置后，这一设备可以很好地复制蜻蜓追踪目标的能力，其原理当然也是蜻蜓捕猎过程中的预判机制。研究人员认为这一技术在该领域具有开创意义。

Wiederman博士说，“根据测试对象的不同，我们会选用相应的摄像系统。我们会对镜头的移动进行测试，以模仿各种类型驾驶员在不同姿势下头部和眼部的移动方式。”若能在实验中使用更大型的无人地面车，就有可能采用更高端（即所需计算能力更高）的算法进行测试。

“我们会测试‘主动视觉系统’，即在闭环条件下，这一移动平台本身是如何影响算法的。借助计算神经学方面的知识，研究小组已经开发出了可以自主选择并追踪移动目标的模型。”

Wiederman博士满怀期待地表示，“我们摁下启动按钮后，接下来就是等着看Clearpath Husky无人地面车自己会选择哪个目标进行追踪。”

CSTMD1神经元

尽管让机器能“看见”移动目标已是一大收获，但是追踪目标的运动轨迹，然后避开其行驶路线，才是对于自动驾驶汽车而言至关重要的。据Wiederman博士称，研究人员发现蜻蜓的这一神经元（已命名为CSTMD1神经元）不仅能预判目标会在何处再次出现，而且可以切换负责追踪该运行轨迹的眼睛，这期间的大脑运转甚至还要在左右脑半球之间进行转换。

eLife期刊上最近发表了一篇有关CSTMD1的研究报告。文章中提到，许多动物都具备在纷繁复杂的环境中侦测移动目标的能力。文中还补充道，这种区分识别的过程其实相当复杂，尤其体现在十分复杂的周边环境中，需要对某个对比度非常微弱的小型目标进行锁定时。

研究认为，蜻蜓具有一种具备“优胜劣汰”选择能力的神经元，可以促使其对某个目标做出选择性关注，而不受其他周边环境的干扰。

更多有关“CSTMD1用于机器人应用研究”的信息，已在2017年7月号的《神经工程学》（Journal of Neural Engineering）期刊上发布。

这一项目为国际合作研究项目，由瑞典研究委员会（the Swedish Research Council）、澳大利亚研究委员会以及瑞典研究和高等教育国际合作基金会（the Swedish Foundation for International Co-operation in Research and Higher Education）共同资助。

A predatory dragonfly’s ability to detect, track and anticipate the escape maneuvers of a juicy target may provide a link to making autonomous driving safer.

Dr. Steven Wiederman, Research Supervisor at the University of Adelaide Medical School in Australia believes that a target-detecting neuron in the dragonfly's tiny brain which anticipates movement, could provide a link to vehicle vision systems. The dragonfly’s visual focus on prey is so great that it can do this even when its target is embedded against a background of “clutter."

To demonstrate the neuron’s potential for safer autonomous mobility applications, a university research team is using an autonomous robot wheeled platform to test sensing techniques derived from the dragonfly. It's part of collaborative research project being conducted by Dr. Wiederman's group and a team at Lund University in Sweden.

The Australian researchers discovered that the target-detecting neuron was able to enhance a dragonfly’s responses in a small focus area just ahead of a moving object being chased. Even if the dragonfly has lost sight of its prey, the focus spread forward over time allows the insect’s brain to predict its likely track and subsequent reappearance to re-establish target acquisition.

A problem facing autonomous vehicles is priority decision making with at least two choices in given traffic situations (or in dragonfly terms, targets). In an interview with Automotive Engineering, Dr. Wiederman explained how the entomological study can help.

“Biological brains have the ability to competitively select one stimulus amidst distracters. We found a neuron in the dragonfly brain that exhibits such selective attention. When presented with two moving targets, the neuron selects just one—sometimes even switching between them mid-trial.

"Sometimes the neuron can ‘lock-on’ to a less salient stimulus," he noted. "We are currently investigating what properties of the target make it the one chosen—is it timing, saliency or trajectory? Is it only attributes of the stimulus, or is the dragonfly choosing the target by some high-order, internal workings in its brain? Finally, when is it appropriate to lock-on, and when is it time to switch to a more salient object?"

By studying this tractable, model system, the researchers hope to gain insight into how more complex brains select stimuli, e.g. a human driving along a road with multiple stimuli. So they're developing models for autonomous, moving platforms based on the dragonfly selection processes.

A unique robotic platform

There are three reasons for using the dragonfly according to Dr. Wiederman, who also heads the Visual Physiology and Neurobotics Laboratory at the Australian Research Council (ARC) Center for Nanoscale Biophotonics. First, it's a fine animal model for electrophysiological recordings; second, it's one of the world’s most effective predators, and third, it exhibits interesting "high-order" processing, e.g. prediction and selection that may not be exhibited by simpler insects, such as the house fly.

Does the dragonfly’s eye have similarities to that of a human? Dr. Wiederman explained that the compound eye has thousands of individual lenses focusing light onto a single retina, while human eyes have one lens focusing light onto a single retina. The dragonfly has less resolution (visual acuity of only ~0.8o), while humans have a central fovea of very high acuity. The dragonfly only has about 10o of binocular overlap and the eyes are too close together, so it must use other techniques for depth perception (e.g. motion parallax).

But Dr. Weiderman noted that there are many similarities between how the underlying neurons process visual information across a diverse range of animal species. By using human psychophysics experiments, his team examines what types of processing are evident in both humans and dragonflies.

The camera-equipped robotic platform is a Clearpath Robotics’ Husky A200 using an open-source serial protocol. Configured at the university, it has been designed to replicate the dragonfly’s target-tracking capability via its predictive pursuit of prey. The researchers believe it to be a technology “first” in such a context.

“We use different camera systems dependent on what we are testing. We test movement of the camera to emulate eye and head movements independent of body directions," he said. The use of a larger ground vehicle provides flexibility to test computationally expensive algorithms.

"We test ‘active vision’—how the moving platform itself affects the algorithms in a closed loop. From the computational neuroscience, the team develops models for autonomous selection and pursuit of a moving target.

"We hit ‘go’ and see what the Clearpath Husky ground vehicle autonomously pursues,” Dr. Wiederman asserted.

The CSTMD1 neuron

While it is one thing for artificial systems to be able to see moving targets, tracing movement so that it can then move out of the way of those things is a significant aspect of self-steering vehicles. The researchers found that the neuron (CSTMD1) in dragonflies not only predicts where a target would reappear but also traced movement from one eye to the other – even across the brain’s hemispheres, Dr. Wiederman reported.

A study of CSTMD1 was recently published in the journal eLife. The article stated that a diverse range of animals have the capability of detecting moving objects within cluttered environments. It added that this discrimination is a complex task, particularly in response to a small target generating very weak contrast as it moves against a highly textured background.

The study refers to the “winner-takes-all” neuron in the dragonfly, which is likely to promote such competitive selection of an individual target whilst ignoring a distraction.

More information on the study of the implementation of CSTMD1 into the robot was published in July 2017 by the Journal of Neural Engineering.

The research project is an international collaboration funded by the Swedish Research Council, the Australian Research Council and the Swedish Foundation for International Co-operation in Research and Higher Education.

Author: Stuart Birch
Source: SAE Automotive Engineering Magazine