Mixed Signature And Dynamic Indexing For Effective Motion Trajectory Representation And Recognition

Motion trajectories contain plentiful information of long term, complex spatiotem-poral motions. Hence, motion trajectories are widely studied for characterizing free form motions of moving objects, e.g. human behavior, robot action and other mo-tion of moving objects. Dynamic motion representation, perception and recognition via trajectory are important for motion analysis. In this way. a flexible descriptor of motion trajectory plays an important role in the performance of motion perception and recognition, especially in3D space. However, in the existing tasks, motion trajectories were usually used in raw data form for motion description, which is not effective for capturing salient features. Shape descriptors and signature of differential invariants are good for local features but global feature is lacking.In this thesis, the attention is focused on the motion trajectory representation, perception and recognition, and how to improve the efficiency and accuracy under free environment. Mixed invariant signature descriptor is proposed with global invariants. The global invariants are capable of capturing the global features that representing the spatial distribution of the motion trajectory. These global features are necessary for distinguishing the confusion among similar motion trajectories to enhance the accuracy of motion recognition. In the motion recognition process, the Continuous Dynamic Time Warping (CDTW) algorithm is modified and used as a matching engine for more accurate. A controllable weight parameter is introduced to fit different tasks under various circumstances.In the second phase, a compact motion description mechanism-dynamic Indexing method with General Shape (IGS) is proposed as a high level descriptor to capture the salient features. In this method, motion trajectories are represented in segment level, which is efficient for fast motion recognition. The trajectory segments correlate to the general shapes which can naturally represent general shape features without predefinition of the parsing rules. Therefore, the motion trajectories represented in segment level can be directly indexed by their general shape sequences. With this indexing method, the motion retrieval and recognition is low in computational load, which is important for the real time applications.The indexing method is extended for multiple object motion recognition in real time applications. To model the parallel temporal relations among the multiple trajec-tories of one motion, a time sequence based model is proposed to characterize them. This method is also invariant for spatial transformation and is effective for real time ap-plications. As the temporal parallel trajectories are indexed in time series, the temporal relations of trajectories can be extracted directly from the indexing database. A max-imum range of the temporal relations is introduced in the training of the time ordered indexing data in datasets, which makes this method robust for motion recognition.A stereo camera system is connected to a computer to capture the trajectory of3D positions of a moving object. The extracted trajectories obtained by this system are used to test the proposed method for motion recognition and perception after the pre-processing of them. The experimental results of motion recognition by mixed signature show the capability of perceiving motion features and distinguish different trajectories with similar shapes. The motion recognition implemented with the indexing method validated the outstanding effectiveness of the proposed method as well as the accuracy. This method is also straight forward and effective for real time applications.