Video Multi-object Tracking Algorithm And Application Based On Deep Fusion Features

Multi-object tracking is a core technology in the field of driverless.As driverless technology gradually enters people’s life,more and more researchers are also participating in the field of multi-object tracking in order to improve multi-object tracking technology.After a long time of technological development,multi-object tracking has begun to be applied in the fields of traffic safety,virtual reality and driverless,and it has also shown great potential in many fields.Nowadays,the related technology of object tracking has been applied to actual life,bringing convenience to people’s lives and improving people’s quality of life.As a basic technology in the field of computer vision,with the development of deep learning and convolutional neural networks,object tracking technology based on deep learning have been widely used in actual life,and has made considerable progress and development.In the traditional multi-object tracking research,the actual traffic environment is usually used as the training scene,and the multi-object tracking algorithm is designed on the typical tracking scene.There are many unavoidable problems in the actual tracking scene.Aiming at various practical problems,this article optimizes the classical tracking model.The multi-object tracking model proposed in this article is improved on the pedestrian multi-object tracking dataset MOT16,and the improved multi-object tracking algorithm is applied to the driverless tram dataset.Tram is a common means of transportation.Compared with other tracking scenarios,tram as a tracking scene not only needs to deal with complex road conditions,but also pay attention to the position of the track line and its own motion state,and use the tracking results in driverless tram to design anti-collision system.According to the various issues mentioned above,the main work of this article can be divided into the following parts:1.For the calculation of target similarity measurement in the registration stage of multi-object tracking model,the traditional similarity measurement method based on pixel features is difficult to effectively represent the similarity between targets due to occlusion or illumination change in complex video scenes.Therefore,in this article,the deep neural network is used to extract the target features and calculate the similarity of the target.At the same time,a variety of relevant methods such as center loss function are used to optimize the feature extractor.In addition,multilevel matching method is also used to improve the accuracy of target matching and get a more effective registration model.2.For the problem of target quantity change in multi-object tracking dataset,this article proposes to use particle filter to realize the generation of trajectory and the judgment of new or dead targets,use the depth feature to optimize the weight of particle filter,maintain and update the tracking trajectory through the particle filter method,and build a more effective trajectory generation model to solve the problems of new or dead targets.3.For the problems of pedestrians and vehicles in the tram scene,based on the improved multi-object tracking model in this article,it adapts to various problems in the tram dataset,and achieves good results in the tram operation data set.While designing the multi-object tracking system,it also provides anti-collision and other functions,and integrates the core multi-object tracking model with multiple functional modules.