Design and evaluation of CNN hardware accelerator designs for in-seat passenger anomaly detection

Passenger distress detection and its notification to the connected ecosystem is of high priority to offer timely help and rescue their lives. Today's Artificial Intelligent (AI) models are capable to extract temporal and spectral features needed to autonomously recognize the passenger state using sensory signals. The work proposes convolutional neural network (CNN) based hardware accelerator design to detect passenger anomaly state, based on the digitized signals acquired from 15 spatially distributed cushion embedded pressure-sensitive-conductive (PSC) sensors. The design was aimed to work successfully on the seated passenger inside the vehicle for three different vehicle conditions including stationary vehicle, vehicle moving on well maintained plain road, and vehicle running on a sub-standard road. The hardware CNN was designed and evaluated for five different configurations by incorporating the reused processing elements that are programmed with the available resources of the targeted FPGA. All the hardware designs were evaluated with respect to power, throughput, resources utilized, and inference delay. The hardware accelerator designed FPGA system exhibited 64.70X, and 22.57X faster inference when compared with other embedded computing implementations including Raspberry Pi, and single core processor in the ZYNQ board.