基于自适应池化的行人检测方法

基于卷积神经网络的行人检测器普遍采用图像识别网络,通常会引起多池化层导致小目标行人特征信息丢失、单一池化方法导致行人局部重要特征信息削弱甚至丢失等,针对以上问题,基于最大值池化和平均值池化方法,提出了一种自适应池化方法,结合通用目标检测器Faster R-CNN,形成了有效的行人检测器,达到增强行人局部重要特征信息、保留小目标行人有效特征信息的目的。对多个公开的行人数据集进行大量实验,结果表明,与传统的卷积神经网络行人检测器相比,所提方法将行人检测漏检率降低了2%～3%,验证了方法的有效性。新方法改进了卷积神经网络结构,在无人驾驶领域具有一定的参考价值。

Pedestrian detection based on adaptive pooling method

Pedestrian detectors based on convolutional neural networks generally adopt image recognition network, which usually causes the following problems:1) multi-pool layers lead to the loss of feature information of small target pedestrian; 2) the single pool method leads to the weakening or even loss of the local important feature information of pedestrians. Therefore, based on the maximum pooling and average pooling methods, an adaptive pooling method is proposed, and combined with the Faster R-CNN, an effective pedestrian detector is formed, so as to enhance the local important feature information of pedestrians and retain the effective feature information of small target pedestrians. Through a large number of experiments on several public pedestrian datasets, the results show that compared with the traditional convolutional neural network pedestrian detector, the proposed method reduces the miss rate by about 2%~3%, which verifies the effectiveness of the method.