结合头部和整体信息的多特征融合行人检测

尺度过小或被遮挡是造成行人检测准确率降低的主要原因。由于行人头部不易被遮挡且其边界框包含的背景干扰较少，对此，该文提出一种结合头部和整体信息的多特征融合行人检测方法。首先，设计了一种具有多层结构的特征金字塔以引入更丰富的特征信息，融合该特征金字塔不同子结构输出的特征图从而为头部检测和整体检测提供有针对性的特征信息。其次，设计了行人整体与头部两个检测分支同时进行检测。然后，模型采用无锚框的方式从特征图中预测中心点、高度及偏移量并分别生成行人头部边界框和整体边界框，从而构成端到端的检测。最后，对非极大值抑制算法进行改进使其能较好地利用行人头部边界框信息。所提算法在CrowdHuman数据集和CityPersons数据集Reasonable子集上的漏检率分别为50.16%和10.1%，在Caltech数据集Reasonable子集上的漏检率为7.73%，实验表明所提算法对遮挡行人的检测效果以及泛化性能与对比算法相比得到一定的提升。

Multi-feature Fusion Pedestrian Detection Combining Head and Overall Information

The decrease in accuracy of pedestrian detection mainly caused by occlusion and too small scale. Since the pedestrian head is not easily occluded and it’s bounding box contains less background interference, a multi-feature fusion pedestrian detection method combines head and overall information is proposed. Firstly, a feature pyramid with multi-layer structure is designed to introduce richer information, feature maps output from different substructures of the feature pyramid are fused to provide targeted information for head and overall detection. Secondly, two branches are designed to perform the detection simultaneously. Then, the model generates pedestrian head and overall bounding boxes respectively from predicted centers, heights and offsets thus constituting end-to-end detection. Finally, non-maximum suppression algorithm is improved to make better use of the pedestrian head information. The experimental results show that the proposed algorithm has 50.16% miss rate on CrowdHuman dataset and 10.1% miss rate on the Reasonable subset of CityPersons dataset, and 7.73% miss rate on the Reasonable subset of Caltech dataset. Experimental results show the detection efficiency and generalization performance of the proposed algorithm are improved compared with the contrast algorithms.