Neuromorphic Data Compression: A Bio-Inspired Paradigm Shift for Real-Time Direct Time-of-Flight Calculation

Direct Time-of-Flight (D-ToF) is predominantly employed for depth-sensing applications. Seeking to increase the spatial resolution without loosing frame rate, the number of readout channels must scale up. Conventional time resolving techniques are based on histogramming. Even by applying the most efficient histogram compression techniques, this solution is not scalable for real-time applications. This is due to the large memory footprint required by high depth resolution and long distance range. This paper presents a paradigm shift in data compression by introducing a bio-inspired neuromorphic approach. It performs time resolving by learning the true ToF from the statistical detection data set. The proposed method relies on a simple Spiking Neural Network (SNN) and operates in real-time. It requires only analog memory, e.g. 12 capacitors are needed for a 12b time-to-digital converter. Considering a 110nm CIS process, the proposed method has a memory footprint of 1980μm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> per pixel. The histogramming method requires 40kb/pixel SRAM memory with a size of 0.601mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, i.e. 303 times larger than the neuromorphic method.