Moving Target Tracking In Airborne Video

Visual tracking is one of the most important research directions in the field of computer vision, mainly due to its wide range of applications in video surveillance, visual navigation, human-computer interaction and video compression and other fields. This paper specifically focuses on moving target tracking in airborne video, and we quote two methods to solve this problem. The first method treats tracking as a binary classification problem of the target and background, using online updated random forests classifier to adapt to the change of the apparent target, to carry the point of robust tracking in a complex environment. The second way uses the Spatio-Temporal Context models to achieve visual tracking, making full use of spatial and temporal relationships between pixels in the local semantic area and pixels in center of the target.In this paper, we use spatio-color histograms to improve the method based on the online random forest classifier and particle filter framework. We use online bagging to simulate the sampling process, the random decision tree is online growing, and random forests method is treated as appearance model of the target in airborne video. In the tracking process the classifiers is updated by new positive and negative samples to adjust to the changes of the target and the surrounding background. In order to increase anti-jamming capability, after obtaining the classification output we calculate spatio-clolor histograms similarity between the particles and the target in the previous frame, and uses the similarity and classification output as the last value of each particle.In this paper, we combine Kalman filter to improve the method based on spatial and temporal semantic model. The model is built with spatial and temporal relationships between pixels in the local context region and pixels in the center of the target. And the model is used to track target in the Bayesian framework. To solve the problem of missing the target after occlusion, we use Kalman filter to estimate the target position, and judge the beginning of the occlusion by the distance between two position. We match the target with the local region near the position estimated by Kalman filter, when the target occurs, the target is located accurately. Experiments show that the anti-blocking ability of our real-time method is greatly improved.