Computer Vision News - July 2019

The next step is about controlling the vehicle to drive autonomously. The control panel has two parts – the algorithm side and the hardware side. The vehicle control unit has to send an electrical signal to the vehicle in order to control the steering and the throttle. This is a full-stack solution from the very beginning to the end for self- driving cars. It is backed up by a data infrastructure built by gathering a huge amount of data from global testing fleets. Before AutoX cars hit the road, they are tested heavily through digital simulation. Jianxiong emphasises how important it is that they make no mistakes. Being a self-driving car, they require 100 per cent accuracy. Not a single compromise is allowed. He is very proud that they have put together such a robust and reliable system. The solution covers many aspects of computer vision. Jianxiong explains: “ For example, for perception, obviously we need object detection. We need semantic segmentation. Not only do we detect the object, but we’re segmenting out at instance-level. We need to do this at a very fast speed. Unlike other applications where you can wait one second to get a picture; for us, we need real-time performance. When the image comes in, we need to finish all the computation immediately. We use a convolutional neural network to do that. ” 43 Computer Vision News AutoX