Deep Learning and Computer Vision based approaches for Airbags deployment analysis by means of user-defined ROIs

Advances in Computer Vision technology paved the way for the development of production pipelines which proved to perform similarly to human operators in terms of accuracy and better then them in terms of time, when applied on repetitive analysis. These performances increase with the introduction of Deep Learning algorithms, which narrow the margin with human level behavior. This Thesis work focuses on Deep Learning approaches for a Computer Vision application in the Passive Safety checks field. In particular, the description of a tool aimed at checking the correct deployment of airbag devices is provided, delving into the Deep Learning-based detection engine. Experiments on the input data to the Machine Learning model are made, focusing on highlighting weaknesses and biases of the current approach. Based on the collected evidences, two proposals for approaches to the problems are introduced. While both are based on Deep Learning implementations, the first solution is an updated and reworked version of the previous methodology. The second approach is far from the logic behind the previous ones and is accompanied by the presentation of integration modules and algorithms that ensure the compatibility with the pre-existing data and application. After a detailed description of the aforesaid alternatives, a test suite implementation is reported and then employed for a fair comparison between the default approach and the introduced ones. The first introduced solution exploits the generalization capacity of the default model, with only few modifications in the field of the training process and output format. On the other hand, the second proposal offers a completely new horizon of solutions aiming at an improvement in the working speed and the simplicity of integration of a simpler Deep Learning task.