基于改进SiamFC的实时人脸跟踪算法

A Real-Time Face-Tracking Algorithm Based on Improved SiamFC

基于现有的人脸跟踪网络存在参数量大、算力高、难以部署到嵌入式平台,无法满足移动设备实时性需求的问题,以SiamFC为基准网络,提出一种基于Two-Way Dense Layer模块改进后的Dense_Block模块。模块在提取特征时具有特征分流、扩大感受野、轻量化网络等优势;为保证人脸跟踪精度且维持实时的在线人脸跟踪速度,通过人脸级联定位搜索策略,先采用浅层的搜索特征和人脸模板特征进行目标人脸初定位,接着对特征响应最大的区域作为深度特征进行人脸重定位,之后,通过NEON指令集优化、知识蒸馏、模型剪枝等方法进一步为人脸跟踪算法加速。实验表明,改进后的SiamFC在部署到RK3288开发板上时,在Accurate、Overlap基本保持不变的情况下,跟踪速度是原SiamFC算法的7.7倍。

Aiming at the problems of large par ameter count,high computing power,difficulty in deploying to embedded platforms,and inability to meet the real-time requirements of mobile devices in existing face tracking networks,a DenseBlock module based on the Two Way Dense Layer module improvement is proposed using SiamFC as the benchmark network.This module has the ad vantages of feature splitting,expanding the receptive field,and lightweight network in extracting features.In or der to ensure the accuracy of face tracking a nd maintain the real-time online face tracking speed,through the face cascading location search strategy,the shallow search features and face template 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,knowled ge distillation,model pruning and other methods are used to further acceler ate the face tracking algorithm.The experim ent shows that when the improved SiamFC is deployed on the RK3288 devel opment 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.