Research And Implementation Of Object Tracking Based On Convoiutional Neural Network

With the development of society, more and more requirement for videosurveillance scene, react accordingly based on video images. However, most of themonitoring mission has a long time, high-risk, high precision characteristics. In orderto more efficiently complete monitoring tasks, using computer vision methodsbecome an important direction to solve the problem. Object tracking in videosurveillance technology is the most important aspect of a good video surveillancesystem, the object must have an efficient and accurate tracking ability. So objecttracking technology in academia has attracted wide attention and research. Objecttracking algorithm in the initial video frame automatically or manually tag is a goodobject tracking, with the subsequent video frames with the goal of change andmovement in real-time and accurate tracking. However, in the context of thecomplicated process of tracking, occlusion, deformation, light and other changes willoccur goal, which gave the track a very difficult task to bring a challenge.Usually feature that tracking algorithm used to describe the ability to objectdifficult to meet the complex diversity goals change tracking process, resulting in theloss of object tracker. In this paper, the difficulties in tracking tasks for learningpresents a tracking algorithm based on depth, the first of tens of thousands of commongoals offline image by convolution depth pre-trained neural network, to get oncommon goals can be expressed from the simple to the complex structural features.Use a pre-training has been structural features, you can get the tracking objectclassification is through online training methods, after which the next object particlefiltering framework for tracking online tracking. Because of the type of imagepre-training is extensive, with a structural feature is used when the object is occludedand changes can still be characterized by the parameter re-training of the object afterthe change again marked. Object tracking methods based on Convolutional NeuralNetwork and particle filter framework solving problems such change occurs goalprovides a good experimental framework. In this paper, in order to better real-time tracking method and accuracy demonstrated in multiple test video sequence showingexcellent tracking capability.