Optimization of a Neural Network for Computer Vision Based Fall Detection with Fixed-Point Arithmetic

摘要

This paper presents an optimized implementation of a neural network for fall detection using a Silicon Retina stereo vision sensor. A Silicon Retina sensor is a bio-inspired optical sensor with special characteristics as it does not capture images, but only detects variations of intensity in a scene. The data processing unit consists of an event-based stereo matcher processed on a field programmable gate array (FPGA), and a neural network that is processed on a digital signal processor (DSP). The initial network used double-precision floating point arithmetic; the optimized version uses fixed-point arithmetic as it should be processed on a low performance embedded system. We focus on the performance optimization techniques for the DSP that have a major impact on the run-time performance of the neural network. In summary, we achieved a speedup of 48 for multiplication, 39.5 for additions, and 194 for the transfer functions and, thus, realized an embedded real-time fall detection system.