In this paper, we present a face tracking method which combines Kalman filter and kernel based mean-shift algorithms. Under face occlusions, the prediction location of Kalman filter can be an alternative for tracking. Also, face scale has been integrated into Kalman filter to improve its performance. Then we extend this tracking method to people counting. Trajectories of faces from tracking are further analyzed to count the number of people in the scenario. Experiments show that the performance of our scale integrated kalman filter has been enhanced; and our tracking method is able to track multiple faces. The results of people counting show a high accuracy under our scenario.