Year of Graduation
Data Organization in Video Surveillance Systems Using Deep Learning Technologies
We consider the organizing of data in surveillance systems by grouping the videos containing identical subjects. The facial regions are detected in each frame and a video stream is split into sequences with one person. The sequences with the same face are grouped using Rank-Order clustering method. Facial features are extracted with the OpenFace, LightenedCNN, VggNet and VggFace2. The experimental study with the YouTubeFaces dataset demonstrated that the most efficient decision is made by matching of L2-normed average embeddings of all frames in a track, which AUC (99.1%) is 2-12% higher than other cases. We found 2105 clusters (36 errors) for LightenedCNN, 2213 clusters (42 errors) for VggNet and 2007 clusters (31 errors) for VGGFace2. The processed information was saved into no-SQL Cassandra database.