Research On Object Detection, Tracking, Interpretation For Optical Measurement Images

The optical tracking and measurement system generates a large number of images during the weapon test.The parameters of weapon's trajectory,pose,and optical radiation characteristic can be obtained through processing these images.These generated parameters are important for the weapon evaluation and the failure analysis.Object detection,tracking and interpretation are the key technology of processing optical measurement images,which first detect the object,then track the object,at last interpret the feature.In order to fulfil the requirements of a variety of test missions,achieve the real-time data processing,and increase the level of automation,we need to study detection,tracking and interpretation methods that are efficient,accurate and adaptive.This paper aims to process the optical measurement images in real-time.In the field of moving object detection,we study background subtraction for the fixed field of view.We also discuss motion detection for the variable field of view.In the field of object tracking,we study region and contour tracking methods based on the incremental learning of the appearance.Furthermore,we explore region tracking approaches based on tracking-learning-detection.In the field of feature interpretation,we improve the feature localization method based on 2D / 3D model.Based on the above mentioned methods,we design and implement a real-time interpretation and estimation system for optical measurement images.This system is successfully applied to various optical measurement experiments to meet the users' requirements of real-time data acquisition and analysis.The main contributions of this thesis include: 1)A method using adaptive background mixture model with spatio-temporal samples.This method evaluates the effect of position and appearance of the adjacent pixels to the decided p ixel,which gives high-quality moving objects detection in the fixed field of view.This method works especially well in the low-quality image sequences,such as noisy or compressed videos;2)A moving object detection method for the near-planar scene based on dense optical flow and homography constraint.This method computes homography constraint of the scene by dense optical flow.The pixels,which don't fit the constraint,are the moving objects.This method gives more accurate result than the frame difference method.Furthermore,it is more efficient than multi-frame method;3)Followed the framework of tracking by detection,a mixture random Na?ve Bayes visual tracker method with online feature selection is proposed.It is based on the spatio pyramid and binary feature of texture and shape.The tracker method adapts to object appearance change,complex scene,illumination change,and partial occlusion;4)An adaptive and accelerated realization of tracking-learning-detection algorithm is proposed.Tracking and detection steps are parallelized to speed up the computation.Moreover,this method adds feature selection and online scale estimation to improve the adaptability of algorithm.It leads to a robust tracking when the object 's appearance and scale changes Furthermore,this method is suited to deal with the problem that object enters and leaves the field of view frequently;5)Followed the framework of tracking by segmentation,a contour tracking based on online learning of Gaussian Mixture Model is proposed.This method can segment and track multicoloured objects precisely and rapidly,becuase the multi-modal is used for describing object appearance;6)A feature localization method based on 2D / 3D model is proposed.It uses shape matching / model tracking to interpret the feature in the real-time image.The method gives accurate and real-time analysis,even if less texture or feature is present in the image.The methods proposed in this thesis can be further developed for other applications,such as video monitor and motion analysis.