Object detection on space-based optical images leveraging machine learning techniques

The expansion of space exploration has led to a soar in the space debris population, increasing the risks of collisions. Addressing this challenge requires advanced space surveillance technologies. Traditional computer vision approaches are unsuitable for real-time applications due to their significant computational demands. Recent progress has been made in ground-based debris detection, thanks to the integration of CNNs. However, to overcome limitations imposed by the atmosphere and other external disturbances, a space-based framework is appealing for spotting fainter objects. This work presents a novel real-time object detection tool designed for space-based applications using machine learning techniques. The absence of labeled datasets for RSOs detection is one of the primary obstacles to AI training, particularly for space-based observations. To tackle this issue, a synthetic ‘.fits’ image generator named SOIG has been developed using photon mapping techniques. The generator produces two types of images. In one instance, it takes into account the sensor’s pointing, which follows the satellite’s attitude. In the other scenario, the sensor’s pointing is considered fixed. Following its training on synthetic images, the subsequent testing phase is conducted through semisynthetic images, which incorporate noise from an actual space-based image. Results demonstrate exceptional performance (mAP50-95 above 90%) for both fixed pointing and rotated and expanded survey pointing images. In the latter case, the tool utilizes sensor attitude information to enhance debris visibility. On the whole, this research wants to contribute to mitigating space collisions and increasing the understanding of machine learning’s potential in space debris detection.