基于自校准卷积网络的行人检测方法

Pedestrian detection based on self-calibration convolution network

针对已有行人检测算法存在的小尺度行人信息描述不充分的问题,提出一种基于自校准卷积网络的行人检测算法。通过将CSP算法的主干网络更换为SCNet自校准卷积网络,有效扩大了网络的感受野范围;将主干网络的低层特征像素信息和高层特征语义信息进行融合,有效促进小尺度行人的检测;对精细的多尺度卷积特征进行多层连接,将行人检测简化为直接的中心和尺度预测任务。实验结果表明,所提算法在数据集CityPersons和Caltech上的平均漏检率为10.97%和4.3%。算法可以增强小尺度行人的信息描述能力,在检测速度方面也有一定的优势。

In view of the problem that the small-scale pedestrian information description of the existing pedestrian detection algorithm is not sufficient,a pedestrian detection algorithm based on self-calibration convolution network was proposed.By replacing the main network of CSP algorithm with SCNet self-calibration convolution network,the receptive field of the network was effectively expanded.The low-level feature pixel information and high-level feature semantic information of the backbone network was fused to effectively promote small-scale pedestrian detection.The fine multi-scale convolution features were connected in multiple layers,and pedestrian detection was simplified as a direct center and scale prediction task.Experimental results show that the proposed algorithm has miss rates of 10.97%and 4.3%on the data-sets CityPersons and Caltech respectively.The algorithm can not only significantly enhance the information description ability of small-scale pedestrians,but also show some advantages in detection speed.