Improved Qualitative Trajectory Calculus for Pair-Activity Analysis

Trajectory provides the most robust feature for activity recognition in far-field surveillance videos, in which increasing attentions have been given to the use of qualitative methods with symbolic rather than real-value features. Qualitative trajectory calculus(QTC) showed a good performance in pair-activity from video. However, QTC and similar works are not good at dealing with noise, since they are all considering short-term features. To deal with the problems mentioned above, two types of long-term features, including sub-trajectory feature and point-trajectory feature, are designed. The sub-trajectory feature is a long-term feature in a coarse granularity, while the point-trajectory feature is a long-term feature in a relatively fine granularity. Using the sub-trajectory feature, a couple of trajectories are segmented into sub-trajectories and enveloping boxes are used to substitute the original sub-trajectory for capturing the major attributes. The point-trajectory feature describes the relationship between a single point in one trajectory and all parts of the other trajectory. The experiments on the human activity classification data demonstrated that our proposed methods are better than the original QTC and previous short-term features.