Machine Learning in Sensors for Collision Avoidance.

Sensors generate a huge amount of data that need to be transferred to a computing device for processing. Such large data transfer takes time and consumes energy. This paper presents a new sensing and computing architecture, referred to as MLIS (Machine Learning in Sensors). MLIS allows a part of machine learning to be done on sensor board thereby dramatically reducing the amount of data transferred to the computing device and hence improving overall system performance and energy efficiency. Using an energy-based probabilistic graphical model, RBM (Restricted Boltzmann Machine), we built a new ADAS (Advanced Driver-Assistance System) computing platform for autonomous driving with phased-array-radar as sensors. A working prototype has been built to provide proof of concept for our new architecture. The prototype is implemented using a TI's mmWave (millimeter Wave) radar board and a Vivado HLS implementation of the RBM on the Xilinx xc7z020-clg400-1 device. Extensive experiments have been carried out using the prototype on realistic scenes on our campus. Experimental results have shown that the proposed architecture can reduce the data to be transferred by a factor of 8 while maintaining 98% accuracy. Based on the experimental settings, we present two case studies that have shown a remarkable reduction in collision probability if applying the new architecture to autonomous vehicles.