Automatic device operation and object tracking based on learning of smooth predictors

Abstract

The disclosure provides an approach for predicting trajectories for real-time capture of video and object tracking, while adhering to smoothness constraints so that predictions are not excessively jittery. In one embodiment, a temporally consistent search and learn (TC-SEARN) algorithm is applied to train a regressor for camera planning. A automatic broadcasting application first receives video input captured by a human-operated camera and another video input captured by a stationary camera with a wide field of view. The automatic broadcasting application extracts feature vectors and pan-tilt-zoom states from the stationary camera input and human-operated camera input, respectively. The automatic broadcasting application further applies the TC-SEARN algorithm to learn a sequential regressor for predicting camera trajectories, based on the extracted feature vectors and pan-tilt-zoom states. The TC-SEARN algorithm itself is able to learn the regressor using a loss function which enables decision trees to reason about spatiotemporal smoothness via an autoregressive function.