Research On Feature-Aided Data Association Algorithm For Multi-Target

With the rapid development in techniques including targets probing and tracking, the methodology of feature-aided data association of multiple targets is exhibiting tremendous advantages. Indeed, this methodology has received extensive attention from researchers at home and abroad, and has become one of the most popular subjects. This dissertation makes intensive investigation concerning relevant techniques of such a methodology. To sum up, the main work and contribution is as follows.1,The dissertation systematically introduces the two essential techniques of multi-target tracking, that is, the technique of filtering estimation and the technique of data association. Subsequently, we make intensive study of the traditional data association algorithms. It turns out that when the density of the clutter is large or the number of the targets is large, the traditional algorithms tend to give wrong association results. Actually, this is because the traditional algorithms only employ the target position information to do the track association, and in the above-mentioned circumstances the measurements of the target position tend to be smearing. In order to solve this problem, we propose the methodology of feature-aided data association with multiple targets.2,To satisfy the requirements of multiple features in the feature-aided data association algorithms, we investigate profoundly the characteristics and the feature extraction methods of the high range resolution profile (HRRP). In addition, by analyzing the characteristics of the ship targets, we propose a new feature selection method. Subsequently, we select partial features using this method to obtain good ones preparing for the subsequent study.3. In fact, the feature-aided data association is proposed to cope with the problem of large measurements error caused by the existence of intensive clutter and multiple targets. In order to reduce the smear of the association probability and improve the tracking accuracy, this dissertation employs the target feature information to modify the joint probability and the traditional position-based statistical distance.Based on the above analysis, the dissertation modifies the traditional algorithms under the frame of the NN algorithm, the PDA algorithm and the GPDA algorithm, respectively.1),Under the frame of NN, we propose the concept of the unified distance. Specifically, we integrate the information provided by the target motion status and target features via the unified distance. This method overcomes the problem of tracking mistakenly of the traditional NN algorithm, and additionally inherits the advantage of low computation burden of the traditional NN. Therefore, the method is of high practical value.2),Under the frame of PDA, we define the feature likelihood function. Specifically, we combine the feature likelihood function with the likelihood function of the position measurements. It is demonstrated that this method improves the joint probability of correct measurement, and also reduce the probability of false measurement. Therefore, the method improves the tracking performance and can satisfy the requirements of multiple targets tracking under certain circumstances.3),Under the frame of GPDA, we estimate the mean and the variance of the current target features by memorizing and calculating the target feature information of the several past instants. In addition, we integrate the target feature information and the target motion status information by using the normal distribution function. By doing this, we optimize the calculation of the cluster probability matrix and improve the tracking accuracy. Compared to the traditional JPDA, the modified GPDA algorithm can reduce the computation burden as well as maintain the association accuracy.