No abstract

Event cameras report local changes of brightness through an asynchronous stream of output events. Events are spatially sparse at pixel locations with little brightness variation. We propose using a visual transformer (ViT) architecture to leverage its ability to process a variable-length input. The input to the ViT consists of events that are accumulated into time bins and spatially separated into non-overlapping subregions called patches. Patches are selected when the number of nonzero pixel locations within a sub-region is above a threshold. We show that by fine-tuning a ViT model on these selected active patches, we can reduce the average number of patches fed into the backbone during the inference by at least 50% with only a minor drop (0.34%) of the classification accuracy on the N-Caltech101 dataset. This reduction translates into a decrease of 51% in Multiply-Accumulate (MAC) operations and an increase of 46% in the inference speed using a server CPU.

No abstract

Much progress has been made in wearable sensors that provide real-time continuous physiological data from non-invasive measurements including heart rate and biofluids such as sweat. This information can potentially be used to identify the health condition of a person by applying machine learning algorithms on the physiological measurements. We present a person identification task that uses machine learning algorithms on a set of biomarkers collected from 30 subjects carrying out a cycling experiment. We compared an SVM and a gated recurrent neural network (RNN) for real-time accuracy using different window sizes of the measured data. Results show that using all biomarkers gave the best results from any of the models. With all biomarkers, the gated RNN model achieved 90% accuracy even in a 30 s time window; and 92.3% accuracy in a 150 s time window. Excluding any of the biomarkers leads to at least 7.4% absolute accuracy drop for the RNN model. The RNN implementation on the Jetson Nano incurs a low latency of 45 ms per inference.