基于多核DSP的运动目标跟踪算法的研究与实现

发布时间：2019-01-05 12:01

【摘要】：随着社会的发展,视频监控系统已走进了千家万户中,智能化是监控系统主要的发展趋势,而在监控系统中实现对运动目标的跟踪是智能化的重要体现。本文将智能视频监控作为研究重点,首先设计了目标跟踪算法的实现平台——基于多核DSP的特色视频监控系统,之后研究了运动目标跟踪算法,最后把算法在多核DSP监控平台上进行了实现。本文的研究重点如下:(1)本文以TMS320DM8168(简称DM8168)多核处理器为核心设计了特色视频监控系统,实现了视频的采集、处理、显示和网络输出等功能。系统的特色之处在于根据实际需求设计了视频流框架,对视频做了TILER变换、添加OSD标志、马赛克拼接等处理,并为了实现视频的网络输出功能,设计了流媒体服务器程序。在设计时通过添加AVS功能提升了系统的稳定性,通过设计PCIe驱动增强了系统的扩展性。(2)在本文中特征提取算法是目标跟踪算法的重要组成部分,本文通过对比多种特征提取算法,选择使用性能与效率兼顾的SURF算法提取目标的特征,并使用多种匹配算法相结合的方式对目标的SURF特征进行匹配。在对目标跟踪算法整体测试时设计了合理的仿真流程,对目标跟踪中的目标形态变化、目标被遮挡、目标影子影响等常见问题,提出了解决方法,并对仿真结果进行了分析。(3)根据DM8168多核监控平台的特点,设计了多个核共同参与的目标跟踪算法实现流程,并根据流程对算法进行了实现。由于算法实现后对实时性要求很高,所以根据DSP的特点对算法进行了深入的优化,使其速度得到了提升。通过对目标跟踪算法的测试验证了实现效果。(4)为了更好的保证算法的性能和速度,把目标跟踪算法中耗时最长的SURF算法在TMS320C6678(简称C6678)平台上进行了实现。在实现时采取划分图片的方式实现对整个任务的划分,使得每个核处理一个子任务。通过对处理结果的分析,证明了这种划分方法的正确性,同时算法的速度也得到了大幅度提升。通过对SURF算法在C6678平台上实现方式的探索为后续目标跟踪算法在DM8168+C6678整个大平台上的实现奠定了基础。
[Abstract]:With the development of society, video surveillance system has entered into thousands of households, intelligent monitoring system is the main trend of development, and in the monitoring system to achieve the tracking of moving targets is an important embodiment of intelligent. In this paper, intelligent video surveillance is taken as the research focus. Firstly, the realization platform of target tracking algorithm is designed, which is based on multi-core DSP video surveillance system, and then the moving target tracking algorithm is studied. Finally, the algorithm is implemented on the multi-core DSP monitoring platform. The research focus of this paper is as follows: (1) this paper designs a special video monitoring system based on TMS320DM8168 (DM8168) multi-core processor, which realizes the functions of video acquisition, processing, display and network output. The characteristic of the system is that the video stream frame is designed according to the actual demand, the video is transformed by TILER, OSD logo is added, mosaic and so on. In order to realize the network output function of the video, the streaming media server program is designed. In the design, the stability of the system is improved by adding AVS function, and the expansibility of the system is enhanced by designing PCIe driver. (2) in this paper, feature extraction algorithm is an important part of target tracking algorithm. In this paper, by comparing various feature extraction algorithms, we select the SURF algorithm, which takes both performance and efficiency into account, to extract the feature of the target, and use a combination of multiple matching algorithms to match the SURF feature of the target. The reasonable simulation flow is designed in the whole test of the target tracking algorithm. Some common problems, such as the change of the target shape, the occlusion of the target, the influence of the shadow of the target, and so on, are proposed. The simulation results are analyzed. (3) according to the characteristics of DM8168 multi-core monitoring platform, the realization flow of target tracking algorithm with multiple cores is designed, and the algorithm is implemented according to the flow chart. Because of the high requirement of real-time performance, the algorithm is optimized deeply according to the characteristics of DSP, and the speed of the algorithm is improved. The result is verified by testing the target tracking algorithm. (4) in order to better guarantee the performance and speed of the algorithm, the SURF algorithm, which takes the longest time in the target tracking algorithm, is implemented on the platform of TMS320C6678 (C6678). In the implementation, the whole task is partitioned by the way of dividing pictures, so that each kernel processes one sub-task. Through the analysis of the processing results, the correctness of the method is proved, and the speed of the algorithm is greatly improved. By exploring the implementation of SURF algorithm on C6678 platform, this paper lays a foundation for the realization of the following target tracking algorithm on the whole platform of DM8168 C6678.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【参考文献】