• GTAV
• object re-identification
• Object-detection
• object-tracking
• syntethic dataset

With the continuous expansion of the city scale, the management of city has become more and more challenging. Thanks to the development of computer vision technology and surveillance network throughout the city, there are many new options for city management, especially in traffic management.
Among them, multi-camera vehicle tracking is one of the important tasks. It aims to track the vehicles over large areas in multiple surveillance camera networks.

The latter is a truly complex task given the high variability of possible scenarios in cities, both urban and suburban. The difficulty is increased considering the various possible weather conditions, vehicles that are very similar to each other, observed from different perspectives and with frequent lighting variations.
Such problems are currently difficult to solve given the unavailability of enough data.
Aiming to fill this gap and thus enable the handling of increasingly challenging scenarios,
this thesis work aims to design, implement and validate a tool for data extraction by exploiting
video game that provides a very realistic simulation of reality and allows totally arbitrary data extraction that can be used in multiple use cases.
Such a tool is subsequently used for the extraction of a synthetic dataset with the objective of fine tuning and improving the performance of a well-known challenge presented by NVIDIA AI City Challenge. The latter is focused entirely in the application of AI to improve the efficiency of operations in city environments.

The previously extracted synthetic dataset has been used to fine-tune a model to perform feature extraction for one of the most important step of multi-camera tracking, the "re-identification" module.It has the objective to distinguish each individual entity, in this case vehicles, uniquely and consequently extract highly discriminative features.

This fine-tuning process has contributed to an improvement in the MOT's main classification metric, the IDF1.This result opens up a wide range of opportunities in the use of the tool made with the possibility of using such data independently or in support of real data.