We show a method of realizing object tracking and image restoration in the dark in which target motion and a reference image are simultaneously estimated using a Bayesian framework. To avoid being trapped in a local minimum in the gradient calculation, a broader search is performed by calculating differences after applying a strong low-pass filter to input images. Deblurring is performed using the motion parameters estimated from the blurred images. As a result, we realized object tracking and image restoration from simulated video images with an SNR of up to -6 dB, and real video images captured in a dark environment of less than 0.05 lx illuminance at the subject surface. In addition, we examined the optimal frame rate for image restoration and we found that a higher frame rate was better under relatively little noise while a lower frame rate was better under much noise.
