A Real-Time Face-Tracking Algorithm Based on Improved SiamFC
Aiming at the problems of large parameter count,high computing power,difficulty in deploying to em-bedded platforms,and inability to meet the real-time requirements of mobile devices in existing face tracking net-works,a DenseBlock module based on the Two Way Dense Layer module improvement is proposed using SiamFC as the benchmark network.This module has the advantages of feature splitting,expanding the receptive field,and light-weight network in extracting features.In order to ensure the accuracy of face tracking and maintain the real-time on-line face tracking speed,through the face cascading location search strategy,the shallow search features and face tem-plate features are first used for the initial target face location,and then the region with the largest feature response is used as the depth feature for face relocation.NEON instruction set optimization,knowledge distillation,model pruning and other methods are used to further accelerate the face tracking algorithm.The experiment shows that when the im-proved SiamFC is deployed on the RK3288 development board,the tracking speed is 7.7 times of that of the original SiamFC algorithm under the condition that Accurate and Overlap keep basically unchanged.
Face trackingSiamFC networkFace cascade locationModel pruningKnowledge distillation
NETL
NSTL
万方数据
