Thesis of Emre Dogan
Estimating human pose and recognizing human activities are important steps in many applications, such as human computer interfaces (HCI), health care, smart conferencing, robotics, security surveillance etc. Despite the ongoing effort in the domain, these tasks remained unsolved in unconstrained and non cooperative environments in particular. Pose estimation and activity recognition face many challenges under these conditions such as occlusion or self occlusion, variations in clothing, background clutter, deformable nature of human body and diversity of human behaviors during activities. Using depth imagery has been a popular solution to address appearance and background related challenges, but it has restricted application area due to its hardware limitations and fails to handle remaining problems.
Specifically, we considered action recognition scenarios where the position of the recording device is not fixed, and consequently require a method which is not affected by the viewpoint. As a second prob- lem, we tackled the human pose estimation task in particular settings where multiple visual sensors are available and allowed to collaborate. In this thesis, we addressed these two related problems separately.
In the first part, we focused on indoor action recognition from videos and we consider complex ac- tivities. To this end, we explored several methodologies and eventually introduced a 3D spatio-temporal representation for a video sequence that is viewpoint independent. More specifically, we captured the movement of the person over time using depth sensor and we encoded it in 3D to represent the performed action with a single structure. A 3D feature descriptor was employed afterwards to build a codebook and classify the actions with the bag-of-words approach.
As for the second part, we concentrated on articulated pose estimation, which is often an intermediate step for activity recognition. Our motivation was to incorporate information from multiple sources and views and fuse them early in the pipeline to overcome the problem of self-occlusion, and eventually obtain robust estimations. To achieve this, we proposed a multi-view flexible mixture of parts model inspired by the classical pictorial structures methodology. In addition to the single-view appearance of the human body and its kinematic priors, we demonstrated that geometrical constraints and appearance- consistency parameters are effective for boosting the coherence between the viewpoints in a multi-view setting.
Both methods that we proposed was evaluated on public benchmarks and showed that the use of view-independent representations and integrating information from multiple viewpoints improves the performance of action recognition and pose estimation tasks, respectively.
Advisor: Atilla Baskurt
Coadvisor: Christian Wolf
Defense date: friday, july 7, 2017