Target point alignment with Gaussian mixture models

Under the remote sensing perception scenario, the sensor field of view alignment is the key step for global information acquisition. However, the model can only utilize position information for alignment, as the remote sensing image has limited resolution and the size of targets is relatively small. And because of inevitable noise and complex background information, false alarms are difficult to eliminate. Moreover, the number of targets that a remote sensor can capture can be scarce compared with the general point registration task. With all these problems, precisely aligning remote sensing images that contain similar targets is extremely difficult. In this paper, we propose Gaussian mixture models with shape and false alarm sensitive strategy to tackle the limited resolution remote sensing image alignment task. We first apply shape context as a structure description to add intrinsic information to our model. Then we introduce a false alarm sensitive strategy to refine the information from the shape context to further enhance the alignment ability of the model. We simulate different target structures and conditions as our evaluation dataset. Thorough experimental results show the effectiveness of our proposed method.