Development of a Compact Neural Network for IR Background Estimation in Wearable Eye Trackers

The majority of wearable eye trackers are based on infrared (IR) oculography. Thus, the presence of intense light background in the IR spectrum, due, for instance, to incandescent ambient lighting (indoor) or sunlight (outdoor), significantly affects the performance of those devices. Here, an IR background identification system is proposed, to be coupled to a low-power eye tracker based on a set of discrete IR photodetectors (PD) to be embedded in smart eyewear. Three approaches were compared: (i) adoption of a dedicated multi-spectral sensor, (ii) use of background acquisitions by means of the same PDs employed for eye tracking, (iii) a frequency-domain analysis of the flicker detection engine integrated in (i). For both the multispectral and PD-based systems, we first evaluated a simple thresholding method and then developed a tiny Feedforward Neural Network (FFNN) to enhance accuracy while maintaining low complexity. The PD-based FFNN achieved the highest overall accuracy (97%), outperforming the multispectral sensor (89%), while also eliminating the need for additional components. Moreover, it required only 28. 5μJ per inference, equivalent to just 0.0014% of a 200mAh battery assuming one ambient light inference per minute for a whole day, making it ideal for power-constrained wearable devices. The preference for the PD-based solution over the multispectral sensor was also supported by the results of the flicker-based approach, which, despite offering cues about indoor/outdoor conditions as well as contextual information such as screen detection, shows limited reliability due to the variability of driver electronics and display technologies.