A scenario based on single-pixel input of recognizing and tracing the critical parts of the moving targets has been proposed. The one-dimensional single-pixel matrix is achieved by bucket measurement by making use of pre-training illumination suitable for the target datasets. The designed recognition network can export the real-time position coordinates of the key components of the moving targets through single-pixel detection. Compared with the general random illumination light field, the results of the pre-training illumination show significant improvement in both recognition efficiency and accuracy. The recognition efficiency can reach 100Hz in the experimental detection of moving model planes. It may have important application prospects in military target tracking and hyperspectral remote sensing.