The Research And Design Of Face Tracking System Base On Improved TLD Algorism

With the development of science and technology and human’s living standards.Face tracking technology has been utilized widely in intelligent city system,elderly care and criminal investigation.And with the increase of aging problems.In elder care areas,service robots with pan-tilt cameras have wild application prospect and huge potential.Because in really scence,monitoring environments are complex and changeable,searching and tracking a specified face correctly in complex environment still have some problems.Although Face tracking and pan-tilt control technology has been welldeveloped in recent years’ research but pan-tilt cameras sold in market still have many problems such as:its accuracy is not high enough,highly effected by outer environment,low automation level which make tracking target stable and accurate become a tough task.To solve these problems above,this thesis put forward a target tracking system based on pan-tilt camera with the background of an elder care service robot.It utilizes tracking algorism to get the position of human’s face,then combines two degrees of freedom pan-tilt control system to control it horizontal and vertical rotation to ensure that the face target always locate at the center of video taken by camera which can reach the best tracking effect.The research contents in this paper are as follow:(1)Four objects tracking algorisms:generative,discriminant,correlation and deep-learning were compared.Amomg those algorisms,deeplearning has high real-time optimization and accuracy.But its algorism model is complex and calculation quantities is huge.Impossible to transplant it to frequently-used hardware platforms to realize the need of elders monitoring robots.So,I choose Tracking-Learning-Detection algorism based on correlation model and online learning mechanism to realize the pan-tilt camera face monitoring’s task need.(2)The Tracking-Learning-Detection algorism based on online learning mechanism is studied.This kind of algorism can update target’s appearance model adaptively.It has good effect when target’s color is similar to the environment or the target disappear shortly in a long-time tracking situation.Analyzing algorism’s three modules:tracking,learning,detection’s principles precisely.And improve algorism’s detection and tracking module such as:decrease detection model’s searching area;combine tracking module with fDDST algorism to improve tracking accuracy.Finally,compare algorisms before and after improvement.Experiments proved that improved aghorism has obvious improvement in tracking accuracy.(3)Study and analyze pan-tilt’s control principle and adaptive control algorism.I also analysised the real-world working requirement.So that we can design the hardware system and control algorism,I designed a low-volume,simple-structure,stable-control pan-tilt hardware system.For pan-tilt control algorism,firstly,the the pan-tilt structure was modelled,control system’s diagram and paraments were establishe.Then with the system identification and simulation in Matlab software,Finally,test the algorism in real pan-tilt control system compare different algorism’s output response to realized the double closed-loop adaptive PID controller’s design.