Doctoral meeting: 'Multiple object visual tracking through deep convolutional networks'
The objective of the doctoral thesis is the development and application of algorithms and deep learning techniques for the tracking of objects in videos using deep convolutional neural networks (CNNs). This is of great interest in many applications such as autonomous vehicles, robotics, or video surveillance, to name a few, as it allows to maintain the identity of the targets over time and obtain their positions and characteristics at any instant.
In particular, the thesis focuses on the tracking of multiple objects without continuous detection information. Special attention is paid to the inference speed of the system, as well as to the computational resources used, as this will allow it to run on embedded GPUs on vehicles equipped with cameras, for real-time operation.
To achieve this, we have developed a CNN with the ability to track multiple objects simultaneously, taking into account the scalability of the architecture to maintain its speed characteristics. These CNNs are able to operate in real-time from live video sources, where occlusions are frequent, scene lighting conditions vary over time, and the appearance of the objects undergoes transformations.