基于多相机特征融合的行人检测算法

Pedestrian Detection Algorithm Based on Multi-Camera Feature Fusion

为了解决多视图行人检测算法对多尺度目标检测效果不佳的问题,提出一种采用膨胀卷积编码进行多视图信息聚合的算法。利用不同膨胀率的膨胀卷积在单层特征层中生成不同尺度的感受野,覆盖目标所有尺度的范围,提高对多尺度目标的检测能力。在Wildtrack数据集上进行仿真实验的结果显示,采用所提算法多目标检测精度最高可达90.7%。

To address the issue of poor detection performance for multi-scale objects in multi-view pedestrian detection algorithms,a dilated encoder method is proposed for aggregating multi-view information.The dilated encoder utilizes dilated convolutions with different dilation rates in a single layer to generate receptive fields of different scales,thus covering the entire scale range of the targets and improving the detection capability for multi-scale objects.Experimental results on the Wildtrack dataset show that the algorithm achieves a maximum multiple object detection accuracy of 90.7%.