基于候选区域选择及深度网络模型的骑车人识别
Cyclist detection based on detection proposals and deep convolutional neural networks

基于骑车人目标识别的骑车人保护系统是保护道路环境中骑车人的重要手段。该文提出了骑车人目标的候选区域选择方法，并结合基于深度卷积神经网络的目标分类与定位方法，实现了骑车人目标的有效识别。候选区域选择方法可分为3部分：骑车人共有显著性区域检测、基于冗余策略的候选区域生成和基于车载视觉几何约束的候选区域选择。在公开的骑车人数据库上进行的对比试验表明：相对于现有的目标候选区域选择及目标识别方法，该方法显著提升了骑车人目标的识别率及识别精度，进而验证了该方法的有效性。
Abstract：Cyclist protection systems based on cyclist detection methods are needed to protect cyclists from road traffic. This paper presents a detection proposal method and a cyclist detection method using deep convolutional neural networks to classify and locate cyclists. The detection proposal method uses cyclist shared salient region detection, redundancy-based detection and geometric constraint-based detection. Tests using a public cyclist dataset show that this method significantly outperforms state-of-the-art detection proposals, which verifies the effectiveness of this method.