Resource efficient machine learning for driver safety

Purpose and goal:

This feasibility study aims to reduce traffic accidents by enabling energy-efficient and low-cost driver monitoring systems based on machine learning. Expected results and effects:

We have shown that convolution operations, which account for 90-95% of the computational cost in the type of machine learning models that can be used in driver monitoring systems, can be made significantly more efficient using pruning on the type of computational platforms used in the automotive industry. This can be exploited to extend the use of deep convolutional networks for driver monitoring as well as other functions that require image recognition.

Approach and implementation:

We worked experimentally by studying convolution operations from a popular image processing network, resnet50. We developed sparse algorithms for convolution and adapted them to the type of processors used in the automotive industry and we compared the performance with similarly adapted non-sparse operations.