一种轻量级网络模型的多尺度热红外行人检测方法

A multi-scale thermal infrared pedestrian detection method based on lightweight network model

现有热红外行人检测算法存在网络参数众多、计算量大和小目标检测效果不佳的问题,针对这些问题,提出了一种轻量级行人目标检测算法。首先,该算法利用轻量级网络Resnext50作为骨干网络,实现检测网络的参数削减以及初始特征提取。其次,在模型中引入特征金字塔模块实现多尺度语义信息融合,并结合多尺度训练策略有效地提高了多尺度目标的检测效果。最后,在数据预处理部分引入图像原色填充策略,该策略有效地防止图像变换尺寸过程中出现的目标失真情况。在自建数据集和公开数据集上实验结果表明,此方法在速度和精度上均取得较好的性能,在自建数据集上mAP达到94.49%,比原SSD高出0.53%,而参数量少了近39%。

Existing thermal infrared pedestrian detection algorithms suffer from extra parameters,large computational efforts and poor detection performance for small targets.To address these problems,this paper proposes a lightweight pedestrian target detection algorithm.First,the algorithm uses the lightweight network Resnext50 as the backbone network to realize parameter reduction and initial feature extraction.Second,the feature pyramid module is introduced to achieve multi-scale semantic information fusion,and combined with a multi-scale training strategy to effectively improve the detection of multiscale targets.Finally,the image primary color filling strategy is introduced in the data preprocessing.It can effectively prevent the target distortion during size transformation for the image.Experimental results on both the self-built and public datasets show that this method achieves better performance in terms of speed and accuracy:it outputs an mAP of 94.49% on the self-built dataset,0.53% higher than that of the original SSD,while retaining nearly 39% less parameters.