When there are hundreds or thousands of cameras producing video streams all day long it is very useful to have an algorithm that analyses such streams instead of a human. Today such technology exists and is called convolutional neural networks for video classification . The downside of such neural networks is that we have a fixed number of cases on which the net is trained which is ok for benchmarking our algorithm on a specific dataset but not for real life applications such as security cameras where we don’t know specifically for which scene the algorithm should give an alert signal. So we need to produce an abstract representation of the video scene (embedding) and to classify it in an unsupervised way .
The first part of this thesis will be research state of the art algorithms for video classification and clustering.
The second part of the project will be devoted to the implementation of such algorithms and testing on real datasets.
Addfor will provide data sets of real-time image sequences obtained in different weather conditions and in non-optimal conditions (flickering and jittering) that represent the common difficulties found on real computer vision tasks.
The final purpose of the thesis will be to define an algorithm which starting from a video sequence will classify similar sequences as belonging to the same cluster in order to give an alert signal when there is an anomalous  sequence belonging to an unknown cluster.
Who we’re looking for
Students that are about to get their master degree in: Mathematics, Physics of Complex Systems.
Skills: Python, advanced math, abstraction skills, exp. w. at least one among TensorFlow / PyTorch.
Duration of this Project: 5-6 months.