Trade-off or appropriate balance between high accuracy on object identification and fast speed of identification process is one of the most challenging problems in the study of pedestrian detection algorithms which is based on convolutional neural network. In this paper, we presented a one-stage pedestrian detection algorithm to optimise the trade-off based on an improved scheme via implying deep network features. Firstly, a novel branch was attached to ResNet-50 backbone network. In comparison to the conventional convolution, a dilated convolution in the branch was used to extract much richer context features. Secondly, a classification regression sub-network with stacking predictors was proposed to locate objects and recognise whether the objects are pedestrians. Finally, a novel loss function was introduced into the scheme to improve our network training method by learning more detailed information regarding pedestrian locations. The proposed scheme in this study demonstrated a competitive missing rate which resulting in 12.90 in the ideal circumstances of accuracy and high speed against the challenging benchmark CityPerson in pedestrian detection.

中文翻译：

---

一种改进的行人检测深度扩张特征提取方案

在基于卷积神经网络的行人检测算法研究中，目标识别的高精度和识别过程的快速速度之间的权衡或适当平衡是最具挑战性的问题之一。在本文中，我们提出了一种单阶段行人检测算法，以通过暗示深度网络特征来优化基于改进方案的权衡。首先，一个新的分支被附加到 ResNet-50 骨干网络。与传统卷积相比，分支中的扩张卷积用于提取更丰富的上下文特征。其次，提出了具有堆叠预测器的分类回归子网络来定位物体并识别物体是否为行人。最后，该方案中引入了一种新的损失函数，通过学习有关行人位置的更详细信息来改进我们的网络训练方法。本研究中提出的方案证明了具有竞争力的缺失率，在准确和高速的理想情况下，与行人检测中具有挑战性的基准 CityPerson 相比，其丢失率为 12.90。