| Title | Real-time Face Tracking in Surveillance Videos on Chips for Valuable Face Capturing |
| Publication Type | Conference Paper |
| Year of Publication | 2020 |
| Authors | Zhao, Qian, Wang, Shengjin |
| Conference Name | 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE) |
| Date Published | oct |
| Keywords | Estimation, face quality estimation, face recognition, face tracking, Human Behavior, Inference algorithms, Metrics, pubcrawl, Real-time Systems, Resiliency, SoC, surveillance, Task Analysis, video surveillance, Videos |
| Abstract | Face capturing is a task to capture and store the "best" face of each person passing by the monitor. To some extent, it is similar to face tracking, but uses a different criterion and requires a valuable (i.e., high-quality and recognizable) face selection procedure. Face capturing systems play a critical role in public security. When deployed on edge devices, it is capable of reducing redundant storage in data center and speeding up retrieval of a certain person. However, high computation complexity and high repetition rate caused by ID switch errors are major challenges. In this paper, we propose a novel solution to constructing a real-time low-repetition face capturing system on chips. First, we propose a two-stage association algorithm for memory-efficient and accurate face tracking. Second, we propose a fast and reliable face quality estimation algorithm for valuable face selection. Our pipeline runs at over 20fps on Hisiv 3559A SoC with a single NNIE device for neural network inference, while achieving over 95% recall and less than 0.4 repetition rate in real world surveillance videos. |
| DOI | 10.1109/ICAICE51518.2020.00060 |
| Citation Key | zhao_real-time_2020 |
Real-time Face Tracking in Surveillance Videos on Chips for Valuable Face Capturing