Extraction and analysis of massive skeletal information from video data of crowded urban locations for understanding implicit gestures of road users

This work explains the possible inferable information from a long-term video acquisition with cameras installed in close proximity to pedestrian movements with an unobstructed view of the entire intersection. The main goal is detecting implicit and explicit gestures and understanding communication and interactions between different types of road users. After explaining the designs of different gesture classification approaches, we relate the qualitative approach with our classification scheme for the extracted skeletons. To this end, a sequence with selected moving entities is selected and compared with the manually annotated video sequence. Results show the limitations of the automated approach and indicate a level of subjectivity in the manual annotation procedure. Subsequently, we discuss possibilities and restrictions of our approach and reflect on the importance of the specific conditions of video acquisitions. Depending on the field of view and distance between installed video cameras and moving vulnerable road users (VRUs), we are able to define the restrictions of our approach. As a result, we are able to define a selection of suitable applications for our approach.