The project goal is to select, implement, customize and test one or few convolutional neural networks (CNN) allowing detection and classification capabilities on edge devices for (outdoor) video surveillance applications.
A smart pan-tilt-zoom (PTZ) camera is like a robot head that rotates to keep a moving object inside the field of view.
The PTZ camera zooms-in as well, to increase far objects resolution making details more visible for identification and recognition.
PTZ tracking is a very difficult task. The problem is that any algorithm at a certain point will drift. It seems unavoidable. What can be done is detecting the drift, re-initiating the target model and resuming the PTZ tracking.
This can be achieved by a proper detection algorithm that needs to be as fast as possible to prevent the target from escaping the field of view permanently.
Deciding when a drift is happening is crucial as well. Further advancements to the overall tracking algorithms may be investigated.
The project activities will include: study of scientific papers; implementation of detection algorithms on embedded platforms; live test execution with PTZ cameras; performance and computational load evaluation; creation of a tagged video set and implementation of automatic tests on a server.
Strong interest in computer vision and machine learning. Basic knowledge (can be acquired/improved during the project): C/C++ and Python programming languages, embedded devices (Linux).
Italian or English
Lunch and Travel