Date of Award
Disparity, Disparity-Augmented Trajectory, Human Activity Recognition, Trajectory
Human activity recognition has been a hot topic for some time. It has several challenges, which makes this task hard and exciting for research. The sparse representation became more popular during the past decade or so. Sparse representation methods represent a video by a set of independent features. The features used in the literature are usually lowlevel features. Trajectories, as middle-level features, capture the motion of the scene, which is discriminant in most cases. Trajectories have also been proven useful for aligning small neighborhoods, before calculating the traditional descriptors. In fact, the trajectory aligned descriptors show better discriminant power than the trajectory shape descriptors proposed in the literature. However, trajectories have not been investigated thoroughly, and their full potential has not been put to the test before this work. This thesis examines trajectories, defined better trajectory shape descriptors and finally it augmented trajectories with disparity information. This thesis formally define three different trajectory extraction methods, namely interest point trajectories (IP), Lucas-Kanade based trajectories (LK), and Farnback optical flow based trajectories (FB). Their discriminant power for human activity recognition task is evaluated. Our tests reveal that LK and FB can produce similar reliable results, although the FB perform a little better in particular scenarios. These experiments demonstrate which method is suitable for the future tests. The thesis also proposes a better trajectory shape descriptor, which is a superset of existing descriptors in the literature. The examination reveals the superior discriminant power of this newly introduced descriptor. Finally, the thesis proposes a method to augment the trajectories with disparity information. Disparity information is relatively easy to extract from a stereo image, and they can capture the 3D structure of the scene. This is the first time that the disparity information fused with trajectories for human activity recognition. To test these ideas, a dataset of 27 activities performed by eleven actors is recorded and hand labelled. The tests demonstrate the discriminant power of trajectories. Namely, the proposed disparity-augmented trajectories improve the discriminant power of traditional dense trajectories by about 3.11%.
Habashi, Pejman, "Trajectory-based Human Action Recognition" (2018). Electronic Theses and Dissertations. 7475.