Research On Target Tracking Algorithm Based On Superpixel Segmentation

Video target tracking technology is a core research topic in the field of computer vision and image processing.It is also a key issue in smart video surveillance,human-computer interaction,and missile guidance.The problem can be described as follows: In the input video scene,adaptively estimating the trajectory of the target object in the subsequent frame according to the given target initial state.The target tracking algorithm can be divided into single target tracking and multi-target tracking according to the number of tracking objects.According to the difference of the apparent model construction methods,it can be divided into discriminative tracking methods and generative tracking methods.At present,the difficulties to be solved in the field of video target tracking include: recognition of targets under similar background conditions,accurate estimation of target dimensions,and large-scale deformation of target appearance.Superpixel segmentation is an important technique for image preprocessing.By segmenting the image into superpixel blocks with certain semantic meanings,the redundant information of the image can be removed,providing intermediate visual features for subsequent image processing,and reducing the complexity of the algorithm processing.This paper combines superpixel segmentation and target tracking,deeply studies the advantages and disadvantages of traditional superpixel tracking algorithms,and proposes two improved target tracking algorithms based on superpixel segmentation,as follows:(1)Aiming at the problem of the drift of superpixel tracking algorithm under the condition of background interference,this paper proposes a target tracking method based on the feature constraint of superpixels.This method constructs a Trimap of the input image by using the state information of the previous frame,and combines the K-means algorithm to initialize the target foreground and background GMMs parameters.Afterwards,the target foreground mask map is extracted by iteratively optimizing the energy function,and a super-pixel feature constraining mechanism is constructed to constrain the super-pixel block confidence value of the target appearance model constructed by Mean Shift clustering,thereby reducing the confidence value of background interference noise pixels..Finally,the Bayesian framework is used to estimate the maximum posterior probability of the state information of the target.Through experimental simulation,it is proved that the target tracking method based on the super pixel feature constraint can alleviate the tracking failure problem in the background interference situation and improve the robustness of the target tracking in the case of background noise interference.(2)Aiming at the time-consuming problem of traditional super-pixel segmentation target tracking algorithm modeling,this paper proposes an online discriminant superpixel tracking method.This method uses a single hidden layer feed-forward neural network combined with an over-limit learning machine algorithm to supervise the super-pixel features in the target foreground-background feature pool to perform supervised learning to construct a target foreground-background superpixel block discrimination network.At the same time,combined with the fast KNN clustering algorithm implemented by k-d tree,the apparent feature space is finely segmented to build a target foreground-background feature dictionary,which replaces the Mean Shift clustering algorithm to quickly construct the target foreground-background appearance model.Finally,the algorithm uses particle filter to estimate the target position information,and introduces the correlation filter at the same time.By extracting the FHOG feature of the image,the target-scale filter is obtained and the target-scale auxiliary estimation is achieved.In this paper,through the experimental simulation,the effectiveness of the tracking effect of the proposed method is verified by quantitative analysis.