Moving Target Trajectory Classification And Identification Of Research

Trajectory classification and recognition of the moving objects are the basic problem of the movement analysis. Its functions are as follows: interpreting what has happened in surveillance scenes, analyzing and recognizing the trajectory activity patterns of the objects in real scenes, classifying them automatically and intelligently. And what will happen is predicted according to the current states of the object, alarms are sent for the abnormal activity in good time. In the end the safety surveillance can be realized. In this thesis, the trajectory classification and recognition are studied in detail. The main contributions are as follows:1. Due to some trajectories are interfered badly by the noises such as wind etc. during the tracking of objects, a novel method that can judge the trajectories's validity is presented. Through analyzing the length, variance of the coordinates and orientation code of the trajectories between two neighbour frames, preprocess them and get the valid ones as samples for further study.2. Using K Means which can automatically cluster trajectories, a new algorithm based on trajectory space similarity distance is presented, and it is applied to classify trajectory. The results are satisfied. Finally, 305 trajectory samples of 6 classes are obtained, and it lays a good foundation for further trajectory recognition.3. A more detailed studying of 3 classic algorithms on HMM is made, some questions in application are pointed, and the corresponding improvement methods are given for them.4. Using HMM to trajectory recognition is introduced. Firstly, aiming at the complex degree of the trajectories, the model are built for every trajectory pattern, and the training samples are used to get the credible parameters of the model, finally, the maximum likelihood probability of test samples are computed to all of the trained model, the maximum value is saved and the corresponding model is the recognition result. Then train and recognize the samples clustered, and recognition rate reach 87.76 % and 94. 19% respectively.