Spiking Neural Networks for Ultra Low Power Event-Based Optical Flow Estimation

Event-based cameras provide an asynchronous, high-temporal-resolution sensing modality that is well suited for low-latency and low-power perception. In parallel, SNNs offer a biologically inspired computational model that naturally aligns with event-driven data and neuromorphic hardware. Together, they represent a promising foundation for energy-efficient motion perception in embedded systems. However, deploying spiking neural networks for dense optical flow estimation on ultra-lowpower hardware remains challenging due to the combined demands of accuracy, memory footprint, recurrent state management, and hardware compatibility. This thesis investigates the feasibility of spiking neural networks for event-based optical flow estimation with a strong focus on deployment-oriented constraints. Starting from a state-of-the-art self-supervised SNN architecture, the work conducts a systematic exploration spanning model architecture, input resolution, temporal windowing, evaluation metrics, and quantization. Particular emphasis is placed on integrating Leaky Integrateand-Fire (LIF) neuron layers into a software-to-hardware toolchain, enabling realistic simulation of memory usage, computational cost, and throughput on PULP-based ultralow-power platforms. A comprehensive trade-off analysis reveals that channel reduction and spatial resolution scaling provide substantial memory and computational savings with limited impact on accuracy, while recurrent spiking layers remain critical for preserving motion direction fidelity. Post-Training Quantization, when combined with careful calibration of LIF neuron states, reduces static memory requirements by approximately 4× with negligible accuracy degradation. Overall, this work demonstrates that spiking neural networks can perform dense event-based optical flow estimation within the stringent constraints of ultra-low-power embedded hardware, provided that architectural design, quantization, and memory hierarchy are jointly considered.