Multi-target Tracking Algorithm Research Based On DSP

Moving target tracking is a very active project in computer vision field, because multi-target tracking detection, identification and tracking technology involved in many subjects such as image processing, pattern recognition and artificial intelligence as well as the change of outside environment and the complexity of multi-target tracking, so tracking multi-target quickly and accurately is the core issue in this subject. This paper researched the correlate methods of multi-target tracking process such as multi-target searching、detection and tracking, improved the basic mean shift algorithm, realized track multi-target continuously, quickly and stably, able to overcome the effect of external environment and noise, also had good robustness.Multi-target tracking mainly includes multi-target capture and multi-target tracking. In the stage of multi-target capture, use median filter technology to pretreat the acquired image, and then according to the difference among multi frame accumulate images, use differential accumulative calculation method to search moving targets and confirm their general position, mark each moving target with image marking technology; confirm all the moving targets, after that use the methods of threshold segmentation and region growing to eliminate errors and wipe out fake targets, obtain the image of multi-target capture which has higher accuracy.This paper uses mean shift algorithm combined with multi wave door technology to track multi-target. Using core calculation formula in mean shift algorithm can calculate center coordinates of candidate model for next frame. But the center coordinates are always decimal, not fit the express mode of image matrix. For the problem of how to round numbers, put forward the concept of drifting similarity, use the integer coordinates which has maximum similarity nearby to replace candidate mode center. In addition, analysis the problem of covering among different targets and objects transient losing in tracking process, use template matching method to recapture the objects that are covered and lose instantly.Test the tracking algorithm in PC and DSP experiment equipments (ICETEK-DM6437), the experimental result shows that, in the status of non-high speed for moving targets, able to track multi targets continuously, has a good real-time, satisfy the requirement of tracking precision in a certain error range.