摘要
时域相关性在视频分析中具有重要作用,但在估算光流时,这一特性却很少被应用。针对这一情况,提出在H-S光流模型基础上引入前向帧,并加入光流时域上的相关性约束构造出一种多帧光流模型以提高光流估算精度。同时,针对光流非线性能量泛函求解异常复杂的情况,提出运用迭代重加权最小二乘法(IRLS)简化模型线性化求解过程。最后,结合所得光流给出一种改进的车辆检测跟踪方案,通过先估算光流后帧差提取,可得到更加精确的车辆运动信息并可避免多个车辆的光流域连接成块。实验结果表明,IRLS法简化求解的多帧模型可同时估算出多个帧间光流并能显著提高光流估算精度,基于此多帧光流的车辆检测跟踪方案的车辆跟踪成功率在86%以上,达到了预期要求。
Despite the fact that time domain correlation plays an important role in analysing video data, this characteristic has hardly been exploited in optical flow estimation. In light of this, we presented to construct a multi-frame optical flow model by introducing the forward frame based on H-S optical flow model and adding the time domain correlation constraint for improving the precision of optical flow estimation. Meanwhile, aiming at the case that it is extremely complicated in solving the nonlinear energy functional of optical flow, we proposed to use the method of iterative reweighted least squares (IRLS) to simplify the process of linearised model solution. At last, in combination with the optical flow derived we offered an improved vehicle detection and tracking scheme. By executing the differential extraction after the optical flow estimation is completed, the scheme could get more accurate vehicle motion information and prevent the problem that the optical flow areas of vehicles may connect to a block. Experimental results showed that the simplified solution of multi-frame model by IRLS can estimate simultaneously multiple inter-frame optical flows and significantly improve the estimation precision, the vehicle detection and tracking scheme based on this multi-frame optical flow reached over 86% tracking success rate and achieved the expected demand as well.
出处
《计算机应用与软件》
CSCD
2015年第9期214-218,共5页
Computer Applications and Software