Abstract
Optical flow is the pattern of apparent motion in a visual scene produced by the relative movement between objects and an observer. Optical flow is used in many engineering applications such as optical odometry. A variety of optical-flow algorithms has been proposed in the past few decades; however, most of these algorithms involve complex computation, making them difficult to be implemented in neuromorphic systems that operate based on neural networks. Interestingly, studies have shown that insect visual systems are able to perform complex optical flow algorithms. Inspired by the classic Reichardt motion detection model proposed for insects, we designed a spatial-Temporal filtering Reichardt (STR) model. This model computes optical flow based on simple filters in the spatial and temporal domains. The STR model is hardware friendly: it does not require time-consuming iteration processes nor computationally intensive multi-layer convolutional networks, which are typical in other optical flow algorithms. We systematically investigate the performance of the STR model with different parameters including: object size, speed, luminance, and filter forms. We also compare the performance of the STR model to the classical Farneback algorithm, and we demonstrate that the STR model is comparable to the classical algorithms while requiring much less computational power.