Deep learning based computer vision for aerial-view street object detection and classification

This project consists of the training of a convolutional neural network based on the YOLO approach for the systematic detection and classification of objects in a street as seen from a drone flying over them, to then be used for an algorithm to predict these objects’ paths and consider possible accidents between them. In the first part, a graphic user interface for dataset creation and validation has been created, with the goal of allowing the creation and validation of other datasets in order to have a custom-made dataset for the goal. The dataset based on is the VisDrone dataset, which consists only of drone-view imagery of streets. After this, using the Darknet framework, several YOLO and tiny-YOLO models have been trained through Google Colab cloud computing services, and then all these have been tested in front of other data and compared to other pre-existing models in the field. Finally, the outputs of the YOLO network is then used in an algorithm to predict possible accidents on streets. This is based on logic-based algorithms for object tracking, linear regression for path prediction, and solver of linear system of equations for the detection of possible collisions, followed by filtering of these. Overall, a YOLO model with a mean average precision of 75% on the development set and 62% on the test set has been trained, and preliminary positive results have come from the accident prediction algorith