基于区域预测和视觉注意计算的快速目标检测

Fast target detection based on area prediction and visual attention computation

提出了一种结合区域预测与视觉注意模型化计算的快速目标检测方法.通过分析图像近似均匀的3个水平子区域的方向特征图之灰度比率,灰度特征图之信息熵和子区域位置,建立了目标区域预测的判定准则.同时,通过优选特征和优化特征图之权重,改进了视觉注意计算模型.对于一幅待检测图像,根据区域预测的判定准则,实现目标区域的快速预测,并利用改进的视觉注意计算模型对目标区域进行视觉注意计算,实现特定目标的快速精确定位.实验结果表明:针对户外场景中的行人目标,与通过整幅图像的视觉注意计算来实现目标检测的传统方法相比较,该检测方法可使检测时间缩短30%,同时还能使检测准确率提高9%.

A new method of fast target detection was proposed with combination of area prediction and modeling computation of visual attention （ VA）. A rule set to predict target area was established based on the analysis of three feature parameters in three horizontal subareas： grayscale rate, entropy and its position. Meanwhile, the VA computational model was improved by optimizing the selected features and optimizing contribution weights of feature maps. Given an image to be detected, the target area was predicted according to the rule set. Then, the VA modeling computation was carried out only in the predicted areas. The experiment results demonstrate that in the task of detecting pedestrian in outdoor scene, the proposed method can reduce the detection time by 30% and enhance the detection accuracy by 9% in comparison with the traditional method.