Ear tracking via Siamese hierarchical refinement network for local active noise control

Abstract

Active noise control (ANC) technology has been applied to reduce unwanted sound in the vehicle cabin. In this paper, a real-time ear tracking system assists ANC performance as the driver’s head moves around. For long-term robust ear tracking, an offline-trained ear detector initializes target area. With precise pre-cropped image patches, a Siamese hierarchical refinement network (SHRNet) builds high-fidelity feature map based on Siamese pyramid branch. Hierarchical feature extraction with lateral refinement makes most use of all levels of feature representation. The offline matching network is trained in an augmented dataset from the self-collected in-vehicle ear database and the ear-labeled McGill face video database. Further, Q-learning is capable of learning a decision-making policy for refining tracking strategy to improve efficiency. Extensive experiment results in various scenes based on NVIDIA Jetson TX2 show the tracker performs at a real-time speed while maintaining a robust performance. In particular, the method achieves AUC score of 67.6% with 26 fps on self-collected in-vehicle ear database.