Moving Objects Tracking Technique On Automatic Dvring Car

Machine vision as a multi-discipline research field, is widely appliedand developed in many fields of our current society, just like industrialdetection, military reconnaissance, human-machine interaction andintelligent car. Particularly for intelligent car, machine vision plays animportant role in understanding the surrounding environment, and theircombination makes the smart car can simulate human perceptionbetter.Vision based moving target tracking on intelligent car is one of thehot research fields. In this paper, vision-based target tracking system hasbeen studied, which includes three core modules: First, vision-based objecttracking algorithm to obtain the target location; then path planning basedon the target location and laser scanner; and the speed and steeringcontrolling of intelligent car. The main works of this paper are asfollowed.Firstly, the intelligent car and controlling system is analyzed, thecontrolling architecture of which is composed of upper and lower controller.The upper level controller is responsible for sensor data acquisition, processing and decision-making, while the under one is responsible for theunder level controlling. They work corporately to ensure system stability.The main sensors are calibrated and co-calibrated for data fusion to improvethe object location precision. For the purpose of automatic controlling of thecar, steering, throttle and brake system need to be modified. This paperdescribes the mechanical, electrical modification and the control strategiesof each system, which is experimentally verified.Next, problems of target detection and tracking algorithms are analyzed.The former one requires a lot of training samples and is hard to adapt totarget deformation, while the latter is difficult to re-initialize when it comesto a fault. So this paper runs detection and tracking in parallel, improvingeach other. Based on TLD (Tracking-Learning-Detection) architecture, thispaper modifies its detector and tracker, using HOG (Histograms of OrientedGradients) feature, NCC (Normalized Cross Correlation) matchingincremental classifier and corner-based optical flow tracking. Then theintegration of the camera and laser radar data improves target locating basedon the joint calibration.Finally, path planning and tracking are analyzed to realize the targettracking. Path planning could be done by the tangent method based on targetlocation and accessible region, which is detected with laser scanner data. Linear Preview with PID controller is commonly used in path tracking, butthe car is always going along a curve instead of a straight line, as what weshow in pose reckon. So it is improved to a curve relative pose preview baseon kinematic model, and the controlled variable of the vehicle steering iscalculated with PD controller.