Visual Saliency-based Video Target Tracking

Target tracking is an important researching direction in computer vision,It has been widely used in security,video surveillance,intelligent transportation,military guidance and other fields.The current video tracking technology is still relatively weak in practical applications,and is faced with many challenging problems such as poor adaptability.Such as the dynamic tracking,camera and object position relatively drastic changes and need to overcome the target flashing,scene in a wide range of noise,light illumination changes lead to the target characteristics to solve occlusion / after losing to quickly locate the problem.It is difficult to effectively deal with the above problems in the existing target tracking algorithms.It has a visual attention mechanism for human visual system;which guides the most interested scene in the area by human eye to avoid a large range of target search problems,to achieve fast target location.This research provides a new video tracking algorithm by simulating human active visual mechanism.First of video frames for rapid visual saliency detection by using saliency detection results to select the most saliency area with human eye gaze point to judge whether there is suspected target area through the fixation intensity in each frame and finally to track objects by the features in the target area.The main work is as follows:(1)This paper presents a new target tracking algorithm framework,we divide it into free search and tracking.The process of free search,we consider the target area by the adjacent frames of the salient object centroid position if the continuous close,then the algorithm into the tracking state.The process of tracking,we keep the state of tracking if the adjacent salient object regions with similar characteristics,then output the trace;if the continuous regional features of several adjacent frames significantly inconsistent are considered goal lost into the free search state.(2)Aim to solve the tracking problem of flyer in Thermal imaging video,it has manly challenges such as target flashing/losing,branches on the ground,the sky clouds interference and obscuring of target,first we perform spectral residual algorithm for the computing of salient value of each pixel,and then by it to reduce the regular texture and influence of illumination change.Algorithm only focus on a few(< 30)pixels,which has high salient value,and the accumulation/separationdegree and average position of these points always denote the centroid position of salient target.Through the position and shape of most salient target between frames can determine and track the target.(3)For the problem of football games in the video,such as target characteristics change,the change of perspective and shade,multiple similar goal etc.,first perform spectral residual algorithm by fast Fourier transformation,and then get saliency map of color image directly.Algorithms only focus on a few pixels which has high salient value,and then find the most salient target by the accumulation/separation degree and average position of these points.Finally through the position and Characteristics of most salient target between frames to determine and track the target.(4)We do comparison experiment between the proposed algorithm and with two popular algorithms,CT(compressed perception)and STC(space-time context learning)algorithm.The experimental results show that the proposed algorithm runs faster,without the human-computer interaction,almost no parameter adjustment,video tracking effect is better than comparison algorithm.Using visual saliency to simulate the human eye gaze in this topic is to guide algorithm for video target tracking.The benefits from that way that the algorithm can automatic track salient target in video scene,without artificial specified initial target;It can effectively cope with losing of goal which is caused by cameras flashing,can quickly find the goal;It can overcome problem such as the change of illumination in the scene,wide range of texture interference,short-term target occlusion.Related video tracking experiments show that the proposed algorithm has unique advantages over classical tracking algorithm,worth further research.