Towards End-to-End Learnable Object Tracking for Video Surveillance

The computer vision problems are of the great practical interest today. My work touches one of the most demanded areas of knowledge today. These are the tracking objects problems. The main objective of this work was to study the issue of a matching detected objects on the video frames. Of the particular interest was the possibility of creating a neural network architecture that would be able to simultaneously detect objects on the video and display their matches on the adjacent frames. In general, not necessary adjacent, but considering adjacent frames was one of the simplifications for the work. Various approaches (methods) to the key components of this task were considered: multiple objects detection, non-maximum suppression, objects matching. In this work, one of the key neural network architectures that was used was Faster RCNN. This architecture is one of the state of the art architectures in objects detection and includes 2 main modules: a module that returns region proposals of the detections and a module that returns probabilities of belonging to a class for the particular region. Weights of the pre-trained ResNet model were also used - a classical architecture used for tuning various tasks in the computer vision. Using Faster RCNN architecture, there was written a baseline that solves the problem. This baseline showed good results on the datasets. There were tested different datasets during the work. Because of the complexity of the work a simplified dataset was drawn by myself. The main analysis was carried out and the metrics of the quality of the algorithms were calculated on this dataset. It turned out that the baseline gives the higher results according with the considered quality metrics. An untrained (deterministic) object matching algorithm was taken as a baseline.