Abstract:

Aiming at the problems of high miss-detection rate and poor real-time performance of small-scale pedestrian target detection in complex road scenes, an enhanced algorithm for small scale pedestrian detection is proposed. The feature extraction network structure and loss function in this algorithm are improved from RFB（Receptive Field Block） network. First, by means of reverse fusion, the deep feature groups shuffled between the channels of multi-scale feature map are merged into the shallow layer at multiple levels, and the improved RFB module and Normalization layer are added at the same time to make full use of the information between the multi-scale feature layers for small-scale pedestrian detection. Second, the loss function is based on intersection ratio and center distance to solve the problem that evaluation index is not equal to regression loss function. The experimental results show that the missing detection rate of overall pedestrian and small-scale pedestrian in the Caltech pedestrian data set is reduced by 4.7 and 9.0 percentage points respectively, and the average detection time of single picture is 36 ms, which is higher than the average level of similar pedestrian detection algorithm.

Key words: small-scale pedestrian detection, complex road scenes, multi-feature fusion, channel Shuffle

摘要：

复杂道路场景中小尺度行人目标检测漏检率高，实时性较差，提出了一种针对小尺度行人检测的增强算法，对RFB（Receptive Field Block）网络从特征提取网络结构及损失函数两方面进行改进：通过反向融合的方式将多尺度特征图通道间Shuffle后的深层特征组多级融合到浅层，并在采用更浅层特征的同时加入改进RFB模块及Normalization层，充分利用多尺度特征层间的信息进行小尺度行人检测。损失函数采用基于交并比和中心点距离解决评测与回归损失函数评价指标不等价问题。实验结果表明，该算法在Caltech行人数据集上总体行人和小尺度行人的漏检率分别降低了4.7个百分点与9.0个百分点，单张图片平均检测时间为36 ms，性能高于同类算法。

关键词: 小尺度行人检测, 复杂道路场景, 多特征融合, 通道Shuffle