Towards Human Intentions Prediction

Accurately predicting human intentions from partially observed actions is critical in various applications. Although deep learning has shown promise in improving accuracy, most approaches ignore contextual information that impacts human behavior. This project aims to identify characteristics that best represent actions and evaluate deep learning architectures that incorporate contextual information in real time. Data from various sources will be considered, and the most predictive features and architectures will be identified. The models' performance will be evaluated using appropriate metrics such as mAP, mean Average Precision and mcAP, mean calibrated Average Precision [1].