Deep Counting Models and their Use for Person Detection

Beschreibung

The focus of this thesis is to use deep learning-based approaches for person detection in real scenarios with challenges like partially occluded or distant persons that appear small along with the lack of availability of training data and annotations. To achieve this, deep counting models (DCM) are proposed as a novel and fundamental mechanism. A DCM is a deep model for counting a specific category of objects. In this thesis, the theoretical aspect of such models is motivated using the Information Bottleneck principle. A systematic approach to train and create deep models for counting using synthetic images is outlined. An extensive analysis of DCMs is done and found to have properties useful for detection. The insights gained are used to evolve various person detection techniques based on the deep counting model.

Autor*in

Sanjukta Ghosh

Themen in »Deep Counting Models and their Use for Person Detection«

Deep Learning Deep Counting Model Person Detection

Stimmen zu »Deep Counting Models and their Use for Person Detection«

Details

Link teilen