Visual saliency with foveated images for fast object detection and recognition in mobile robots using low-power embedded GPUs

This paper presents a visual saliency algorithm for fast object detection and recognition in mobile robots using low power graphics processing units (GPUs), based on human vision foveation. The step of image foveation enables the use of small images, which leads to a much reduced number of computations in deep convolutional neural networks and consequent increase in frame-rate. We demonstrate how using a simple foveated downsampling method, we can maintain a detection-recognition performance level similar to the level at larger image resolutions, even when transforming from 416x416 to 128x128 pixels, for a small high acuity region of the image, which can lead to a 4× speed up in frame rates, maintaining a relatively stable mean Average Precision. The visual saliency algorithm is evaluated on the Stanford drone dataset and our own experimental drone dataset.