针对道路监控视频提出一种基于视觉注意模型的关键帧提取算法。首先采用自顶向下的方法,通过运动检测获取运动目标,以车牌和车辆最佳清晰度位置作为注意度评价标准,提取运动目标位置显著度;然后在运动目标内部采用自底向上的方法,提取运动目标的运动方向和强度显著度;接着用一种简单有效的车辆位置优先的自适应线性混合模式合成视觉注意度,并在时间方向上生成最终的视觉注意度曲线;最后求出视觉注意度曲线的导数曲线,自适应滤波处理后,在正值到负值变化的零交叉点中选取显著度最高的图像作为关键帧。实验结果表明,本文算法提取的关键帧不但包括了所有经过监控的车辆最佳或接近最佳清晰度的位置,而且还能包括道路停车、超速和逆向行驶等各种交通事件,符合交通观察者的视觉特性,同时也有利于进一步对关键帧进行车辆静态特征的提取,以形成交通视频的特征数据库。

A key frame extraction algorithm based on the visual attention model is proposed for lane surveillance video. First, the top-down method is used to detect moving objects whose position saliency is decided by the clearest position of license plates and vehicles. Then, within the moving objects, the bottom-up method is used to calculate the moving orientation and moving intensity saliency of these moving objects. Next, the visual attention curve is fused by a simple adaptive linear mode. Last a derivative curve is generated, from which the frame with the most salient value in those zero-crossing points from the positive to the negative on the derivative curve is selected as key frame. Experiments show that the key frames extracted by the proposed algorithm not only include the optimal or suboptimal positions of all passed vehicles, but also include on-street parking, speeding, reverse driving, and other traffic incidents. The results are consistent with the traffic observers’ visual perception and conducive to the extraction of vehicle static features to form the traffic video features database.