Abstract: In order to solve the problem of target detection in road traffic scenes, this paper proposes a target detection method based on EdgeBoxes algorithm and deep learning fusion. Using the Fast R-CNN, a deep learning target classification algorithm, combined with VOC format non-motor vehicle data samples, the problem of target detection is factored into the classification of bicycle and evbike. It uses EdgeBoxes algorithm to extract the object proposals of the samples, then to build moderate regions of interest and enter the network together with the sample for iterative training. At the same time, regularization and fine tuning strategies are introduced to optimize the network, reducing network complexity and avoiding over-fitting. After network training the non-motor vehicle target detection model is obtained, a new sample test is performed on the model and the test result is analyzed. Compared with the traditional method, the detection method based on EdgeBoxes algorithm and optimized Fast R-CNN method improves the detection accuracy slightly in road traffic scenes detection, and the computational complexity is significantly reduced and the detection speed is nearly doubled.

Key words: target detection, EdgeBoxes, deep learning, Fast R-CNN

摘要： 针对道路交通场景目标检测问题，提出采用EdgeBoxes算法和深度学习融合的非机动车辆目标检测方法，利用深度学习目标分类算法Fast R-CNN，结合VOC格式的非机动车辆数据样本，把道路交通场景中的目标检测问题实化为自行车（bicycle）和电动车（evbike）的分类问题。利用EdgeBoxes算法提取样本的目标建议构建适量的感兴趣区域，和样本一起输入网络进行迭代训练，同时引入正则化思想和微调策略进行网络优化，降低网络复杂度并避免过拟合现象；网络训练后得到非机动车辆目标检测模型，对模型进行新样本测试并分析测试效果。在道路交通场景目标检测中，基于EdgeBoxes算法和优化Fast R-CNN融合的方法与传统方法相比，检测准确度稍有提高，运算量明显降低，检测速度加快近一倍。

关键词: 目标检测, EdgeBoxes, 深度学习, Fast R-CNN