Fast Target Tracking Based on Adaptive Smooth-scale Particle Filtering

To address failures caused by changes in the target scale during visual tracking, we adopt particle filtering to track the target scale, and propose a fast tracking method based on adaptive smooth-scale particle filtering. First, to enhance the ability to identify the same target at different candidate scales, we construct a probability distribution feature based on a directional gradient. Then, to improve the accuracy of the scale estimation, we use prior knowledge to optimize the suggested density distribution function in the particle filtering algorithm. Finally, we use a reasonable scale constraint function to perform a second optimization of the estimated target scale. In addition, to reduce the interference caused by large-scale scaling, we set an appropriate threshold to adaptively update the scale of the kernel correlation filter when using it to predict the position, which effectively improves the accuracy of the position prediction. The experimental results show that the proposed scale-estimation algorithm performs better than several typical scale-estimation algorithms in terms of scale estimation and real-time performance.