小型嵌入式平台实时目标跟踪算法的研究与实现

针对高实时性要求、低计算能力的小型嵌入式平台的应用背景,本文提出一种低时间复杂度、高鲁棒性的目标跟踪算法。首先,构建基于时空上下文贝叶斯概率模型的跟踪算法架构,然后提出低时间复杂度的灰度特征尺度池策略实现尺度自适应更新,最后利用基于置信图最大似然概率的目标模型更新策略来提高抗遮挡性能。利用基准数据集OTB2013对本文算法进行测试,跟踪精度为58.9%,成功率为51.3%,优于时间复杂度相近的STC(Spatio-Temporal Context)和CSK(Circulant Structure with Kernels)算法。搭建以DSP为核心的小型目标跟踪平台对算法进行测试,可实现对视场中目标的实时稳定跟踪。当目标波门为64×64 Pixel时,稳定跟踪帧率可达42 frame/s,能够满足实时性和工程实用性的应用需求。

Research and Implementation of Real-Time Target Tracking Algorithm for Small Embedded Platform

Aiming at high real-time requirements in engineering but limited computing power small embedded platform,this paper proposed a low time complexity and high robustness of target tracking algorithm.Firstly,the infrastructure of tracking algorithm was established based on the spatio-temporal context Bayesian probability model.Then,a low time complexity gray scale characteristic scale pool strategy was proposed to realize scale adaptive updating.Finally,the target model updating strategy based on the maximum likelihood probability of confidence graph was proposed to improve the anti-occlusion performance.The benchmark data set OTB2013 was used to test the comprehensive performance of the algorithm,with a tracking accuracy of 58.9%and a success rate of 51.3%,which was superior to STC(Spatio-Temporal Context)and CSK(Circulant Structure with Kernels)algorithms with similar time complexity.The proposed algorithm was tested on a small target tracking platform with DSP as the core,which could realize the stable tracking of video sequences.When the 64×64 target gate was taken,the tracking frame rate was 42 frames per second,which could meet the requirements of real-time performance and engineering practicality.