Research On Multi-feature Fusion Tracking Based On Feature Reliability

Nowadays,with the rapid development of artificial intelligence technology,intelligent technology affects human life all the time.Computer vision is one of the core field of artificial intelligence.Among them,the development of target detection and tracking technology has deep influence on our daily life.In recent years,the technology of target tracking has made great progress and has been widely used in various fields.However,due to the complexity and changeability of the application scenes,such as: the deformation,the illumination change and the occlusion of the target,the tracking performance is not ideal.To address the existing problems at the present stage,this paper proposes a multi-feature fusion tracking algorithm based on feature reliability.In this paper,moving target detection and extraction based on optical flow technology are studied.First,the optical flow field of two adjacent frames is calculated by using the optical flow algorithm.Then the two-dimensional optical flow field is converted into a three-dimensional RGB image by visualization method.Then color clustering method is used to attribute the pixels of the RGB image into two categories: target and background.Then,the target pixel is extracted by threshold segmentation algorithm.Finally,morphological expansion and corrosion operations are used to refine the detection results.For target tracking,this paper first uses the FHOG feature and color histogram feature of target to train the correlation filtering models to produce two responses.Then the two responses were fused into appearance feature responses according to their feature reliabilities.Then,using the optical flow field calculated above,the motion features of the target are obtained through vector synthesis,and the motion features are trained by the correlation filter template to obtain the motion feature response.Finally,according to the feature reliability,the appearance feature response and motion feature response are fused into the final response.Among them,the position with the largest response is the optimal position of the target,and on this basis,the optimal scale is obtained.In order to ensure the accuracy and robustness of tracking,a method of calculating characteristic reliability is proposed in this paper.Based on the bayesian statistical method,the pixels in the Ho G map and optical map are classified into two categories.And two kinds of distribution reliability are produced.Then the proposed algorithm divide the pixels into target regions and backgrounds according to the probability as target pixels.The probability is an expression of the reliability of the features,and it is used as a feature fusion weights.The final tracking experiment is carried out on the OTB100 dataset.The experimental results show that the tracking effect of the proposed algorithm in a variety of scenarios shows good robustness,and the tracking effectiveness of the proposed algorithm is demonstrated by analysis and comparison with some algorithms in recent years.