融合检测和跟踪的实时人脸跟踪

Real-time face tracking based on detecting and tracking

目的在实时人脸跟踪过程中,因光照变化、目标被遮挡以及跟踪时间长等因素,导致的误差累积都会影响系统的整体性能。针对这些问题,提出一种融合检测和跟踪技术的方法,其中包含了检测、控制和跟踪3个模块(简称DCT)。方法在检测模块中,利用AdaBoost算法提取人脸的相关信息,并将信息传递给跟踪模块进行跟踪处理;在跟踪模块中,采用在线随机蕨和SURF(speeded up robust features)算法对目标进行跟踪。同时,在每次检测到目标之后,会通过控制模块对当前跟踪目标准确性进行判断。结果选取国际标准数据组并与LBP+Camshift+Kalman滤波算法、SEMI算法、TLD(tracking-learning-detection)算法比较,实验结果表明,DCT方法在目标发生尺度较大变化、目标遮挡、旋转、形变以及光照发生变化时都具有良好的跟踪识别效果,DCT方法识别准确率在95%以上,平均误识别率和漏识别率分别为0.86%和0.78%。结论 DCT方法具有消除误差累积,跟踪失败后自动恢复等特点,同时可以消除环境中光照、遮挡和仿射变换的影响并满足系统跟踪的实时性要求,运用于视频人脸跟踪系统中能够提高系统的实时性及鲁棒性。

Objective With the high-speed development of computer technology and the need for video monitoring applications, face detection and tracking are gradually becoming a research hotspot and are widely used in public security, intelligent video surveillance, authentication, and others. Considering factors such as variation in lighting conditions, target obscuration, and long-term tracking, tracking human faces produces various errors that reduce the performance of the entire system. To resolve these problems, this paper presents an approach with a combination of detection and tracking technologies, and involves the three modules of detection, control and tracking （DCT） . Method In the control modules, human face features are detected though haar features, and calculated with integral images. This reduces computational complexity and enhances system speed. The AdaBoost algorithm is then used to extract a face＇s information in the detection module. Next, the information is transferred to the tracking module. The face target is tracked by using online random fern and Speeded Up Robust Features （SURF） algorithms in the tracking module. To improve tracking speed, 2-BitBP features are described in the random fern. Then, the P-N learning method is used to quickly locate similar objects. Similar targets and matching confidence of the target are determined based on the NCC algorithm. The similar target with the largest confidence value is considered the tracking target. Whereas, the control module is used to judge accuracy of the current tracking after each target is detected. The interactive unit in the control module transfers the testing results into the tracking modular for subsequent tracking. Three filters in the check unit are designed to filter check results in terms of scale size, target position, and similarity. This balances the performance of detection module and tracking module and avoids tracking failure because running the system for a long time and interference of similar targets. Result By using international standard data sets, the experimental results compared with LBP ＋ Camshift ＋ Kalman filter algorithm, the SEMI-supervised learning algorithm, and the Tracking-Learning-Detection （TLD） algorithm, show that the DCT approach proposed in this paper has good tracking effect on large scale target changes, occlusion, rotation, deformation and illumination changes. This effect can be achieved at the level with identification accuracy of over 95% , the average error recognition rate of 0. 86% and missing recognition rate of 0. 78%. On tracking accuracy, when light varies greatly, the experiments of the target center offset show that the DCT approach can still track the target position accurately. In terms of tracking performance, the average running time of this approach is 46. 25 ms, and its effective rate is 98.25%. Conclusion DCT can eliminate error accumulation, and automaticaUy detect the target once tracking fails. Moreover, it can reduce the influence of environment lighting, obscuration, and affine transformation and meet performance requirements of real-time tracking in the augmented reality system. This approach can improve real-time performance and robustness of the system when applied to a video face tracking system.