Low cost and power CNN/deep learning solution for automated driving

Abstract

Automated driving functions, like highway driving and parking assist, are increasingly getting deployed in high-end cars with the ultimate goal of realizing self-driving car using Deep learning techniques like convolution neural network (CNN). For mass-market deployment, the embedded solution is required to address the right cost and performance envelope along with security and safety. In the case of automated driving, one of the key functionality is “finding drivable free space”, which is addressed using deep learning techniques like CNN. These CNN networks pose huge computing requirements in terms of hundreds of GOPS/TOPS (Giga or Tera operations per second), which seems beyond the capability of today's embedded SoC. This paper covers various techniques consisting of fixed-point conversion, sparse multiplication, fusing of layers and network pruning, for tailoring on the embedded solution. These techniques are implemented on the device by means of optimized Deep learning library for inference. The paper concludes by demonstrating the results of a CNN network running in real time on TI's TDA2X embedded platform producing a high-quality drivable space output for automated driving.