Silicon Designs for Core Autonomy Algorithms (SiDeCAA)

"Affordable, vehicle-mounted cameras are an effective way to access high quality image data from a vehicle's surroundings. However, this data contains no information about vehicles, cyclists, pedestrians, traffic signs or road markings. We have to deduce this higher level information from the images using software, which must run in real time to keep up with the camera output.

Trained, deep learning algorithms, have shown remarkable, near human, ability in analysing image content and this can be used to drive autonomous system behaviours.

There are limitations however, both in terms of their accuracy and their performance, which restricts their deployment in autonomous vehicles.

First, by only analysing single frames from the video stream, we are losing relevant information from any preceding frames and in a traffic environment of largely known fixed objects moving in mostly, well defined paths, this time based information is vital and must be captured to improve scene understanding for self driving cars.

Second, running these large networks on existing processors does not meet the restricted power and performance criteria required by car makers.

Our collaboration seeks to address address both these limitations and produces new silicon chip designs that will generate revenue for UK Plc.

By working with Alex Kendall from the University of Cambridge Machine Intelligence Lab, we will develop a new, time coherent, semantic image segmentation algorithm, which will show improved object classification accuracy and speed. Then Myrtle will realize this as a new custom processor design which we will demonstrate in an existing test car, benchmark it in a test environment and then seek to license to gobal car makers"