Research On Object Tracking Algorithm Based On Particle Filter And Spatial Reasoning

As a hot issue in the field of computer vision object tracking is widely used inthe field of human-computer interaction, intelligent traffic and visual navigation, andit also involves the interaction of many disciplines, including image processing,artificial intelligence and pattern recognition. Video object tracking can be understoodas the extraction and analysis of characteristics of a specific object in successiveframes, through the feature matching and ultimately lock and track the object.Among the current research in the field of object tracking, we generally extract acharacteristic to describe the goal of representation, however, for a large number ofdata, a feature extraction represents often not efficient, and in the object trackingprocess, often encounter a variety of confounding factors, such as shelter, theinfluence of illumination, distortion and mobile imaging, this leads to a feature not agood solution for these confounding factors. Additionally, it only simple tracking theobject in the past, but the research using space direction of motion into objecttracking is relatively less. In order to object features are extracted efficiently, in thispaper, we adopt a multi-feature fusion methods,put SURF feature and HOG featurefused and consist features forests, based on adaptive weights allocation to assignfeature weights and calculate. At the same time, object tracking is actually todetermine the position of the object, analysis the object trajectory and so on in aparticular space environment, spatial reasoning contains a study of the spatialorientation of the relationship between moving objects. For the question of changingfor the spatial location in the object tracking process, this paper introduces spatialreasoning knowledge, we construct a360degrees spatial orientation model which hasa strong ability to express, and combine particle filter to realize comparative andmatching of images, introduce the object tracking algorithm design process and stepsin detail, and give the experimental analysis in two environments--indoor and outdoor.The main research contents of our paper are as follows:1. We fully summarize the current research on object tracking technology, andanalyze the main difficulties from the current object tracking. All in all, the analysisand research on this content is the solid foundation and warm-up for paving ourobject tracking work.2. Then, we specially introduce the local SURF feature and global HOG feature.Because HOG is a dense feature, its matching effect for the characteristics of theobject from small area is better than SURF feature. However, the SURF feature issparse feature, which has very good effect in calculating time and processing rotationor movement. We propose the improvement for traditional vocabulary tree model, anduse the k-means method to establish the dynamic vocabulary tree independent forSURF feature and HOG feature for each particle in the process of tracking, thenself-applicable determine the fusion feature vocabulary tree weights, make sure thatthe good feature can be assigned to a large weight, on the other hand, the givenweight is very small, eventually form the spatial representation model using featureforest for object extraction work, express the characteristics of the object imageeffectively, and obtain efficient object features.3. Through the spatial representation for the object image and particle images,combining the flexibility of particle filter, we put forward the direction relationsmodel of spatial reasoning with360degrees based on the projection-based model, notonly realize the similarity between the image matching features, determining theposition of object in the next frame, but also introduce the changes of object duringobject tracking process, ultimately realize the tracking process of a specific goal.In this paper, we extract features of images in adaptive weights allocation bybuilding feature forest. It can extract features efficiently, at the same time, this featureextraction method have a good resistance for confounding factors, for example, light,occlusion, and so on.This paper firstly introduce spatial reasoning reasonableknowledge to the field of object tracking, combine with particle filter, the algorithmdetermines the object position in the next frame, at the same time can introduce theposition change of the object, achieve a real time, accurate tracking of a specific object.