Neuromorphic machine learning for fast object recognition using dynamic vision sensors

如果你无意中走到一辆自动驾驶的无人驾驶汽车前面，你会想确保它能够在撞到你之前紧急停车。要做到这一点，它的反应时间需要在几毫秒的量级。依靠安装在车辆上的传统的基于帧的摄像机在这个时间尺度上做出响应是不现实的，因为每个帧之间的时间大约为20毫秒。更有问题的是，它们在光线不足、大雨或大雪或高对比度阳光/阴影条件下的表现不佳。动态视觉传感器（DVS）提供了一种新的相机技术，其操作方式类似于人类视网膜，一旦检测到活动的变化，在几微秒内，在弱光和高对比度下，就从每个图像像素发送信号。在自动驾驶汽车等应用中使用这些传感器需要开发新的人工智能，特别是机器学习技术。该项目旨在开发神经形态技术，其灵感来自大脑如何处理来自视网膜的信息，该技术将在DVS检测到视觉事件时迅速识别和响应视觉事件，并为最苛刻的应用提供快速目标检测和识别。

If you step out inadvertently in front of an autonomous, driverless vehicle, you will want to be sure that that it will be capable of making an emergency stop before hitting you. To do so, its reaction times will need to be of the order of a few milliseconds. Relying on conventional, frame-based cameras mounted on the vehicle to respond in this time scale is unrealistic, given the 20 milliseconds or so between each frame. Even more problematic is their poor performance in poor lighting, in heavy rain or snow, or in high contrast sun/shade conditions. Dynamic Vision Sensors (DVS) offer a new camera technology which operates like the human retina, sending out signals from each image pixel as soon as it detects a change in activity, within a few microseconds, in poor light and under high contrast. Using these sensors in applications such as autonomous vehicles requires the development of new Artificial Intelligence, in particular Machine Learning technology. This project aims at developing neuromorphic technology, inspired by how the brain processes information from the retina, which will recognise and respond to visual events as quickly as the DVS detects them, and provide fast object detection and recognition necessary for the most demanding applications.