Multi-camera parking lot vehicle re-identification with a siamese network

This thesis aims to propose a vehicle re-identification algorithm that is using the information of several cameras placed around a parking lot to track multiple vehicles over time. To this end, a siamese deep neural network is designed and trained to model the similarity between two images. In addition, several metrics are proposed to quantify the quality of the information contained in a multi-view vehicle representation. Furthermore, a parking lot dataset is recorded to test the algorithm in real conditions. The experiments show on one hand the improvement brought by the number of cameras and the multi-view vehicle representation and on the other hand, the importance of selecting the information according to a quality criterion to improve the re-identification performance. It also identifies the limitations of the proposed approach.