Research On Video Based Human Objecet Tracking And Recognition

Video based target tracking and recognition technology, is one of the main research directions of computer vision, and also is the base and key technology of applications, such as smart surveillance, human computer interaction, terrain navigation, intelligent video retrieval, smart city, safe city and so on, so it has great values in the theoretical research and practical application. However, there are more complex changes with the surroundings and targets in videos recorded naturally and uncontrollably, which brings many challenges for all kinds of targets tracking and identification in them. How to implement high robust real-time multi-target tracking and identification technology which suitable to various complex scenarios has become the industry urgently needs. In view of the above understanding, basing on a series of outstanding achievements before, and being inspired by the perfect mechanisms found in the target tracking and recognition system of human beings, the further researches have been done mainly on the tracking and identifying for human objects in videos, and some innovative methods with excellent high robustness and effects have been proposed in this thesis.The main works and contributions accomplished in this thesis are summarized as following:1. Aiming to address the problem that the features of the targets would to be invalid due to being sheltered, made-up or lost with time and so on, a novel one sample based autonomous online features learning and updating algorithm(FLUA, for short) is proposed in this thesis, in which once the target is located by the known features, its new features would be learned and updated by muti-verification, such as the matching of values, distributions and motion consistency between feature points, and then participate in the subsequent tracking and recognition process immediately. The experimental results show that, without a lot of samples and training in advance, even only with one sample of the target, this algorithm can also achieve very good effect, e.g., the tracking and recognition rate can be improved greatly, and it can satisfy real-time system.2. Aiming to resist the adverse factors in the video face tracking, such as head posture or facial expression changes, partial face shelter, make-up, deliberately disguise and so on, a human body fuzzy tracking based video face tracking and capturing algorithm(B-FTC, for short) is proposed in this thesis, in which once the target body is tracked by the muti-verification, e.g., the matching of features and motion consistency, the target face would be located and captured by its location and correlation to the body, accompanied with the online feature learning and updating mechanism to cope with the gradual appearance changes of the target body. Experimental results prove that, this algorithm has excellent immunity to facial cosmetic, camouflage, the changes of the head viewing angle or facial expression and so on, and has perfect effect in face tracking and ownership classification, e.g., with the natural surveillance and the face performance videos, its tracking and capturing rate is over 90% and the correct rate is perfect 100%.3. Aiming to address the problem that most of facial images being tracked in the surveillance videos are various facial fragment due to the natural changes of the characters facial expressions and head views, which results in the failure of the face recognition algorithms based on frontal or nearly frontal calm faces, this thesis puts forward an adaptive facial space and expression double weighted based Video-to-Video face recognition algorithm(SEDW-2VFR, for short), in which the reference facial patches is classified into different expression sub-sets by muti-verification with values, distributions and the motion transform errors between feature points, and then the feature projection matrixes and feature matrixes are calculated with the sub-sets. By then, the target facial patch is identified on-line with adaptive weighted 2D-PCA. Experiments show that, the proposed algorithm has strong robustness to head posture and expression changes and has good recognition effect, e.g., its correct recognition rate is over 90% in the natural surveillance videos.4. Aiming to address the problem that most of existing gait recognition methods preset too more conditions demanding, and their gait presentation, extraction and matching processes are relative complex, resulting in large amount of calculation and poor recognition effect, a limbs and gait period double distinguished feature asynchronous extraction and synchronous weighted fusion based gait recognition algorithm(FAESWF-GR, for short) is put forward in this thesis, in which the reference gait cycle features of each limbs are asynchronous extracted and classified into different time-slice sub-sets by their cycle length firstly, and the gait feature projection matrixes and gait feature matrixes are calculated with the sub-sets secondly, and by then, the target limbs gait cycle features are extracted and identified on-line with synchronous weighted fusion based 2D-PCA. Experiments show that, the algorithm has the stronger robustness and the higher correct recognition rate.