Research Of Video Target Tracking Methods Based On Particle Filter

With the introduction of intellectualized concepts, such as intelligent transportation, safe city, peace government, etc., intelligent video monitoring technology gradually becomes one of the fundamental technologies of modern society which features informatization and networking, among which the issue of effective tracking of the moving objects in video sequence has become a hot research topic. The tracking algorithm based on Bayesian’s reasoning became the main method of visual tracking technology, its specific ideas is converted the target tracking into the Bayesian estimation. Making a careful observation and recruiting the maximum posteriori probability of the objective by acquiring new target when the prior probability of target state is known. But in the practical process for visual tracking, the characteristics of the posterior probability distribution is nonlinear, non-Gauss, multi-modal. And therefore researchers adopt the standard Particle Filter (Particle Filter, PF) method to solve the target tracking problem, which became the main method in this filed. There are three results of visual tracking technology:particle degeneracy, reliable observation model and accurate motion model. Normally, it is impossible to set the accurate motion model through2D image list, therefore researchers remove the particle degeneracy problem by establishing the reliable observation model. But a lot of problems in reality are restricting the development of these technologies, such as the sudden change of illumination, attitude change, partial occlusion or complete occlusion, quick movement and maneuvering target, which effect the robustness and accuracy of target tracking. Sponsored by the Ministry of Education Doctoral Program (20106201110003), in order to improve the accuracy and robustness video target tracking system in different complex environments, this article conducts systematic and in-depth research on particle filtering algorithm and processing fusing characteristics of target, obtaining the following achievements:1.Optimal mechanism of particle filter method suggested distribution function is establishedThe guiding theory of standard Particle Filter (Particle Filter,PF) method is to solve the post probability density of system state with the summation approximation, so that the state matrix corresponding integrated operation can be effectively avoided. What works is the sample mean method to the system post state estimation. In standard PF algorithm, for the sake of convenience, the one simple step transition probability of the system is adopted as the optimal suggested distribution function, due to that the lack of information on the latest observation model corrective action can lead to the weights recession. Based on the association between observation information and weights decline,this paper proposes the correction thought and implementation steps towards the weight decline, updates the measurement of sampling particles with the integral kalman Filter and improved string iterative no trace kalman filtering method, puts forward two methods of improving Particle Filter method, and applies the corresponding method to moving target tracking system in video sequence, effectively improving the tracking accuracy. Because a universal concept is adopted in this algorithm, the distribution function optimization mechanism suggested in this paper is applicable to most methods targeting optimal operation, open to adaptability.2. An adaptive filtering method under unknown noise is put forward Most of the current filtering methods are suitable in the circumstance where the system noise is precisely known, but in practice, the statistical characteristics of the system noise cannot be predicted, especially in the complicated scenes where the sudden light change, camouflage and the like take place, which may change the statistical characteristics of system noise, in addition, and the changable noise can directly cause the mismatching of system model, thereby reducing the overall system tracking accuracy, even resulting in tracking failure. Therefore, in order to solve the problem mentioned above, this paper puts forward the adaptive filtering method under unknown noise, which estimates the system noise statistical properties in a real-time manner by use of the Sage-Husa estimator, and introducing the No Trace Kalman Filter Method into the Sage-Husa estimator measurement updating, which effectively avoids the divergent phenomenon of the noise statistical properties’estimation.3.The particle adaptive sampling and filtering divergence inhibition mechanism is establishedUsually, PF method of filtering accuracy is in proportion to sample particle number, but not in all cases, a larger sample particle number is desirable, which may slows down the calculation speed, hindering the real-time. Therefore, this paper carries on an in-depth analysis on the effect of the systematic observation income residual on the sampling efficiency, introduces the deviation of estimated and predicted information, which can self-adapt sampling particle population online. Therefore, in the case of fewer sample particles, the matching of the system model can be enhanced and a high sampling accuracy can be maintained while the number of sampling particles is being reduced. At the same time, the latest measurement information is adopted to decide the filtering divergence trend, and the attenuation memory factor is introduced to effectively suppress the divergence, better the first ensuring the semidefinite and definite of the system noise variance array, effectively enhancing the accuracy and robustness of video target tracking system based on this method.4. Video multi-features adaptive fusion mechanism is achievedAlong with the development of multi-source information fusion technology, multi-features fusing technology gains a unique advantageous edge in the enhancement of the tracking accuracy and robustness, but the current fusion method is mainly to adopt multiplicative fusion and additive fusion. Multiplicative fusion can sharpen the target probability distribution, enhancing the identification of the probability density, but inhibits the multimodality of the state distribution, strengthening the system noise; Additive fusion is characterized by using feature weights weighted summation method to work out the final right value adjustment factor of target system. Although the disturbance of noise on the system estimation is lessened, the credibility of the tracking performance does not make a huge difference. To surmount these shortcomings, this paper proposes the method of differentiating credibility of different characteristics grounded on the principle of maximum information entropy, achieving a video multi-features adaptive fusion mechanism. The method is applied to vehicle tracking of different scenarios on the highway, effectively improving the robustness and accuracy of target tracking in complex environments.The theory of PF algorithm proposed in this paper can be applied on a broader scope, effectively enhancing the robustness and accuracy of target tracking in intelligent monitoring system in different complicated environments.