People detection and tracking using deep learning based approaches

Abstract

The primary objective of this work is to develop an efficient and rapid proposal for detecting and tracking persons based on images captured by a mobile robot. This will be achieved through the integration of deep learning-based detectors and state-of-the-art tracking algorithms. Pre-trained detectors on the COCO dataset will be evaluated to identify the most effective ones for human detection, and tracking capabilities will subsequently be added through cutting-edge algorithms. The effec- tiveness of the solution will be measured using specialized datasets and specific performance metrics. Additionally, a new dataset, ROBOT_7, will be created, designed to reflect the operational scenarios of the mo- bile robot PeopleBot. An extensive experimentation has been carried out in order to identify the best combination of detector and tracking algorithm for this application. As conclusion, we propose a specific com- bination of detector and tracking algorithm that achieves high levels of F1-score and CLEAR MOT performance achieving rates of frames per second good enough for a real-time performance.