Visual Tracking of Materials and Astronaut Operation Action Recognition in Space Station Cargo Spacecraft Cabins

The automatic recognition of cargo pick-and-place actions in spacecraft cabins is important for improving the intelligence level of in-orbit cargo management. This study proposes an integrated visual recognition framework based on improved YOLOv11n, DeepSORT tracking, and SVM-based trajectory classification. To address the challenges of densely stacked cargo, local occlusion, and complex cabin backgrounds, WeightConv and the convolution and attention fusion module (CAFM) are introduced into the YOLOv11n detector to enhance cargo feature representation. Based on the detection results, DeepSORT is used to associate cargo targets across video frames and extract continuous motion trajectories. The trajectory descriptors are then classified by an SVM into three typical operation states: “pick,” “place,” and “not picked.” On the self-constructed cabin dataset, the improved detector obtains a precision of 97.3%. The tracking module obtains an MOTA of 97.1% and an IDF1 of 95.1%, while the trajectory-based SVM achieves an overall classification accuracy of 92.0%. Experimental results demonstrate that the proposed framework provides reliable visual evidence for automating the recording of typical cargo pick-and-place operations in spacecraft cabins, offering viable technical support for enhancing in-orbit cargo management efficiency.