Interactive Tracking, Prediction, and Behavior Learning of Pedestrians in Dense Crowd

The ability to automatically recognize human motions and behaviors is a key skill for autonomous machines to exhibit to interact intelligently with a human-inhabited environment. The capabilities autonomous machines should have include computing the motion trajectory of each pedestrian in a crowd, predicting his or her position in the near future, and analyzing the personality characteristics of the pedestrian. Such techniques are frequently used for collision-free robot navigation, data-driven crowd simulation, and crowd surveillance applications. However, prior methods for these problems have been restricted to low-density or sparse crowds where the pedestrian movement is modeled using simple motion models.;In this thesis, we present several interactive algorithms to extract pedestrian trajectories from videos in dense crowds. Our approach combines different pedestrian motion models with particle tracking and mixture models and can obtain an average of 20% improvement in accuracy in medium-density crowds over prior work. We compute the pedestrian dynamics from these trajectories using Bayesian learning techniques and combine them with global methods for long-term pedestrian prediction in densely crowded settings. Finally, we combine these techniques with Personality Trait Theory to automatically classify the dynamic behavior or the personality of a pedestrian based on his or her movements in a crowded scene. The resulting algorithms are robust and can handle sparse and noisy motion trajectories. We demonstrate the benefits of our long-term prediction and behavior classification methods in dense crowds and highlight the benefits over prior techniques.;We highlight the performance of our novel algorithms on three different applications. The first application is interactive data-driven crowd simulation, which includes crowd replication as well as the combination of pedestrian behaviors from different videos. Secondly, we combine the prediction scheme with proxemic characteristics from psychology and use them to perform socially-aware navigation. Finally, we present novel techniques for anomaly detection in low-to medium-density crowd videos using trajectory-level behavior learning.