基于改进门控上下文聚合网络的雾天实时行人检测方法

A Real-Time Pedestrian Detection Method Based on Improved Gated Context Aggregation Network in Foggy Weather

针对自动驾驶车辆在雾天情况下易将行人误检和漏检的问题,提出一种基于改进GCANet除雾网络和CenterNet检测网络相结合、有效进行雾天行人识别的联合检测方法。该方法在GCANet中引入结合底层细节和全局结构的复合损失函数,优化除雾图的结构细节及图像质量;并将改进的GCANet除雾算法应用于检测算法的训练图像预处理中,最后送入CenterNet网络训练。试验结果显示,本文提出的方法在合成雾天数据集Foggy Citypersons上的平均对数漏检率MR-2值达到9.65,在真实雾天数据集RTTS上的平均精度AP50值达到86.11,降低了雾天场景下行人的漏检和误检情况,有效提升了检测网络在雾天条件下的泛化能力。

For the problem that pedestrians are easy to be misdetected or missed by autonomous vehicles in foggy weather,a joint detection method based on improved GCANet defogging network and CenterNet detection network is proposed to effectively detect pedestrians in foggy weather.Firstly,a composite loss function combining the underlying details and global structure is introduced into GCANet to optimize the structural details and image quality of the defogging map.Then,the improved GCANet defogging algorithm is applied to the training image enhancement preprocessing step of the detection algorithm,which is finally sent to the CenterNet for training.The test results show that the average logarithmic missed detection rate of the proposed method on the synthetic Foggy Citypersons dataset reaches 9.65,and the average accuracy value on the real foggy RTTS dataset reaches 86.11.The proposed method reduces the missed detection and false detection of pedestrians in foggy weather,which effectively improves the generalization ability of the detection network in foggy weather.