针对全卷积神经网络对单帧红外图像行人检测计算量大、检测率较低等问题, 提出了一种改进的 LeNet-7系统对红外图像行人检测的方法。该系统包含 3个卷积层、3个池化层, 通过错误率最小的试选法确定每层参数, 以波士顿大学建立的 BU-TIV数据库训练系统。首先, 以俄亥俄州立大学建立的 OTCBVS和 Terravic Motion IR Database红外数据库作为测试图像；然后, 采用自适应阈值的垂直和水平投影法得到感兴趣区域( regions of interest, ROI)；最后, 将得到的 ROI输入训练好的系统进行测试。3个测试集检测实验表明, 本文方法具有良好的识别能力, 与不同实验方法相比, 本文方法能有效提高检测率。

We proposed an improved method of pedestrian detection in infrared images based on the LeNet-7 system, to address the problems of large computation and low detection rates in traditional methods based on a full convolution neural network. The system consists of three convolution layers and three pooling layers. The trail selection method with the smallest error rate is used to determine the parameters of each layer, while the BU-TIV database, established by Boston University, is used to train the system. Firstly, the Object Tracking and Classification in and Beyond the Visible Spectrum(OTCBVS) and Terravic Motion IR Database, established by Ohio State University, are used to test images. Then, the region of interest (ROI) is obtained by vertical and horizontal projection with adaptive thresholds. Finally, the ROI is input into the trained system for testing. Experiments on three test sets demonstrate that the proposed method has good recognition ability. Compared with different experimental methods, the proposed method can effectively improve the detection rate.