The increasing demand for high-resolution hyperspectral images from nano and microsatellites conflicts with the strict bandwidth constraints for downlink transmission. A possible approach to mitigate this problem consists in reducing the amount of data to transmit to ground through on-board processing of hyperspectral images. In this paper, we propose a custom Convolutional Neural Network (CNN) deployed for a nanosatellite payload to select images eligible for transmission to ground, called CloudScout. The latter is installed on the Hyperscout-2, in the frame of the Phisat-1 ESA mission, which exploits a hyperspectral camera to classify cloud-covered images and clear ones. The images transmitted to ground are those that present less than 70% of cloudiness in a frame. We train and test the network against an extracted dataset from the Sentinel-2 mission, which was appropriately pre-processed to emulate the Hyperscout-2 hyperspectral sensor. On the test set we achieve 92% of accuracy with 1% of False Positives (FP). The Phisat-1 mission will start in 2020 and will operate for about 6 months. It represents the first in-orbit demonstration of Deep Neural Network (DNN) for data processing on the edge. The innovation aspect of our work concerns not only cloud detection but in general low power, low latency, and embedded applications. Our work should enable a new era of edge applications and enhance remote sensing applications directly on-board satellite.
19 March 2021 in Journal of Real-Time Image Processing
Smoke detection represents a critical task for avoiding large scale fire disaster in industrial environment and cities. Including intelligent video-based techniques in existing camera infrastructure enables faster response time if compared to traditional analog smoke detectors. In this work presents a hybrid approach to assess the rapid and precise identification of smoke in a video sequence. The algorithm combines a traditional feature detector based on Kalman filtering and motion detection, and a lightweight shallow convolutional neural network. This technique allows the automatic selection of specific regions of interest within the image by the generation of bounding boxes for gray colored moving objects. In the final step the convolutional neural network verifies the actual presence of smoke in the proposed regions of interest. The algorithm provides also an alarm generator that can trigger an alarm signal if the smoke is persistent in a time window of 3 s. The proposed technique has been compared to the state of the art methods available in literature by using several videos of public and non-public dataset showing an improvement in the metrics. Finally, we developed a portable solution for embedded systems and evaluated its performance for the Raspberry Pi 3 and the Nvidia Jetson Nano.
Individual spacecraft manual navigation by human operators from ground station is expected to be an emerging problem as the number of spacecraft for space exploration increases. Hence, as an attempt to reduce the burden to control multiple spacecraft, future missions will employ smart spacecraft able to navigate and operate autonomously. Recently, image-based optical navigation systems have proved to be promising solutions for inexpensive autonomous navigation. In this paper, we propose a robust image processing pipeline for estimating the center and radius of planets and moons in an image taken by an on-board camera. Our custom image pre-processing pipeline is tailored for resource-constrained applications, as it features a computationally simple processing flow with a limited memory footprint. The core of the proposed pipeline is a best-fitting model based on the RANSAC algorithm that is able to handle images corrupted with Gaussian noise, image distortions, and frame drops. We report processing time, pixel-level error of estimated body center and radius and the effect of noise on estimated body parameters for a dataset of synthetic images.
Modern digital cameras use specific arrangement of Color Filter Array to sample light wavelength corresponding to visible colors. The most common Color Filter Array is the Bayer filter that samples only one color per pixel. To recover the full resolution image, an interpolation algorithm can be used. This process is called demosaicing and it is one of the first processing stages of a digital imaging pipeline. We introduce a novel data-driven model for demosaicing that takes into account the different requirements for reconstruction of the image Luma and Chrominance channels. The final model is a parallel composition of two reconstruction networks with individual architecture and trained with distinct loss functions. In order to solve the overfitting problem, we prepared a dataset that contains groups of patches that share common chromatic and spectral characteristics. We reported the reconstruction error on noise-free images and measured the effect of random noise and quantization noise in the demosaicing reconstruction. To test our model performance, we implemented the network on NVIDIA Jetson Nano, obtaining an end-to-end running time of less than one second for a full frame 12 MPixel image.