Research On Surveillance Video-based Attendance Record And Stranger Validation

Nowadays video surveillance systems play an important role in public safety and home protection, as they appear everywhere. It becomes a trend to make video surveillance more intelligent, which brings us a huge amount of merits.This article focuses on indoor video surveillance-based face recognition and proposes a system through the comparisons of various algorithms. The system realizes real-time processing and achieves high accuracy using only a few video data.The main work and contributions in this paper are as follows:1. We implement a video surveillance system with the functions of attendance and security, which consists of three models, including moving object detection, face detection and face quality assessment, face recognition. Extensive experiments on either ChokePoint dataset or real scence prove the high accuracy and good practicality of our system.2. We propose a modified ViBe algorithm for moving object detection. By adding video stability judgement and background rectification, our algorithm achieves stable background detection, reduces the noisy pixels and solves the shawdow problem. The comparisons with other methods on two published datasets demonstrate that our algorithm has a great performance on precision rate, recall rate and F-score.3. We evaluate different face keypoints detection and face quality assessment approaches. The result indicates that the LBF-based keypoints detection can achieve real-time application with high precision and small memory storage. The face quality assessment based on learning to rank can take into consideration various conditions of face images and thus has robustness.4. We conduct experiments of different features(i.e., DeepID feature, LBP, LPQ, and HDLBP) and the different classifiers(i.e., SVM, KNN, and random forest) on ChokePoint dataset. Finally, DeepID feature and SVM classifier are adopted for their high performance. Finally, our system shows great generability and robustness on unconstrained video faces.5. We also present the detailed implementation of our system and the reasons of designing each module. To consider the performance of both attendance record and stranger validation, we propose a two-step approach which combines face identification and face verification, and finally improve both the ability of real-time processing and the system accuracy.