适用于智能视频监控的立体匹配算法

Stereo matching algorithm for intelligent video surveillance

在智能视频监控系统中,针对双目立体匹配实时性的要求以及由于光照和噪声大容易引起误匹配的问题,提出一种基于改进AD-Census和两级自适应引导滤波的立体匹配算法。采用设有噪声容限并且引入梯度信息的AD-Census算法计算匹配代价,减小噪声和光照的影响;采用两级自适应引导滤波算法快速进行代价聚合,提高算法的实时性;提出一种多步提精方法进行视差优化,有效对遮挡区域和误匹配区域进行深度值估计;利用GPU(graphics processing unit)加速使该算法基本满足实时性要求。在Middlebury标准测试平台下对该算法进行测试,实验结果表明,该算法在光照变化和噪声大的环境下均能得到准确的视差,基本满足实时性的要求。

In the intelligent video surveillance system,a binocular stereo matching algorithm based on improved AD-Census and guided filtering was proposed to meet the real-time requirement of stereo matching and to solve the problem of mismatching due to illumination and noise.The AD-Census algorithm with noise margin and gradient information was used to calculate the matching cost and reduce the influence of noise and illumination.The two-stage adaptive guided filtering algorithm was used to realize the cost aggregation quickly and improve the real-time performance.A multi-step refinement method was proposed to optimize the dispa-rity map,the depth value was estimated for the occlusion area and the mismatched region.The algorithm satisfies the real-time requirement by GPU acceleration.The algorithm was tested under the Middlebury standard test platform.Experimental results show that the proposed algorithm can obtain accurate disparity map in the environments with different illumination and noise,it can basically meet the requirements of real-time.