Research And Implementation Of Object Tracking Algorithm Based On MobileNet

Object Tracking is an important task in computer vision domain and widely applied to robot,video monitoring and military weapon,which is aimed to predict the trajectory of moving object and the size of one.That is to say,for every frame in the video,the confirmed object must be found from the background,namely,the object tracking algorithm outputs a bounding box,including location and size.In this paper,we propose a high real-time object tracking algorithm,the combination of MobileNet,Ensemble Kalman Filter and Correlation Filter,named as MCEK algorithm by initials of every model.The framework of the tracking algorithm consists of a feature extractor based on the MobileNet,a trajectory predictor based on Ensemble Kalman Filter,online learning mechanism based on adaptive multi-branch Correlation Filters and a fusion.The MobileNet based feature extractor outputs multi-layer features which are respectively the input of multiple correlation filters to achieve online learning of parameter,and an adaptive weight scheme is integrated into the framework to fuse these independent multi-branch correlation filters.Due to the low computational complexity and spatial complexity,MobileNet help real-time tracking much.The role of the trajectory predictor based on the Ensemble Kalman Filter is to optimize the prediction results of the above multiple correlation filters in this process.The following is the main work of this paper :First,this paper adopts a lightweight MobileNet,which makes the tracking algorithm achieve better real-time performance under limited computational performance and ensures the online learning of parameter.And different layers of convolution features can represent different information,and achieve a better feature representation by multi-layer feature extraction to cope with various challenges in target tracking.Secondly,in this paper,because the direct linear estimation of the variance matrix in the solution of the variance matrix in the Kalman Filter easily results in the performance instability and even the numerical divergence,an Ensemle Kalman Filter is proposed.The result of the single filter and the motion estimation value are respectively synthesized to obtain the final tracking result.Thirdly,we conducted experiments to verify the performance of the algorithms proposed in this paper.Algorithm experiments show that the proposed algorithm has better accuracy than other well-known algorithms.Moreover,the algorithm proposed in this paper can achieve better results on robot mobile platforms or embedded systems.This paper introduces the MCEK object tracking algorithm.The feature of the object area are extracted by the MobileNet and it is input to the respective correlation filters,and the prediction results are coupled,and the Ensemble Kalman Filter optimizes the output of each correlation filter.Based on the above work,we have fully verified the feasibility and performance of the algorithm.