Search

An intelligent, person-following nanosatellite is under development for in-cabin astronaut assistance in the China Space Station. It is named the Intelligent Formation Personal Satellite (IFPS). The satellite weighs 2.0 kg and is shaped as a sphere of diameter 230 mm. Fans and MEMS flywheels are used for its position and attitude maneuvering as inside the Space Station cabin it is a weightless and standard atmospheric pressure environment. The RGB-D camera and IMU based visual-inertial SLAM method is used to support its localization and navigation. The on-board information processing and computing hardware primarily consists of an embedded AI microprocessor and an FPGA. The satellite is designed to fly autonomously and follow the designated astronaut to offer immediate assistance. Thus, efficient and robust astronaut visual tracking is the most important prerequisite for supporting its basic person-following operating mode. We achieved this by further improving our previously proposed tracking algorithm that consists of a deep convolution neural network (DCNN)-based detection module and a probabilistic-model-based tracking module. The DCNN in the detection module was further improved through optimizations of lightweight network architecture design, parameters model compression and inference acceleration. While maintaining the originally high detection accuracy, the DCNN was optimized significantly in terms of memory, computation and power consumption to quite meet the engineering constraints in the development of IFPS. The complete pipeline of the astronaut visual tracking algorithm was also designed and implemented in the embedded AI microprocessor for online tracking application. Experimental results demonstrated the effectiveness of the proposed efficient and robust astronaut detection and tracking algorithm. • An in-cabin nanosatellite for astronaut assistance is designed. Deep-learning-based astronaut visual tracking algorithm is proposed for supporting its astronaut-following operating mode. • An efficient DCNN was developed through lightweight network architecture design, parameters model compression and inference acceleration. • The DCNN provides high detection accuracy, and its memory requirements, computation burden and power consumption are engineering affordable. • The deep-learning-based tracking algorithm runs efficiently on embedded hardware. Robust detection and tracking performance is verified by experiments. [ABSTRACT FROM AUTHOR]