一种面向自动驾驶路况的目标检测算法

为了对常见的行人和车辆进行检测,采用自行标注的数据集,通过基于faster region-based convolutional neural network （RCNN）框架的算法进行调参与优化. 主干网络采用轻量化网络MobileNetv2,在原生锚框的基础上,区域建议网络（RPN）部分增加2个面积尺度,检测部分使用感兴趣区域（ROI）Align结构,减少特征图映射和均分过程中的误差. 实验结果表明：使用faster RCNN目标检测网络,可以有效完成行人和车辆的检测任务,整体效果良好.

A target detection algorithm for autonomous driving scenarios

In order to detect common pedestrians and vehicles, a self-annotated dataset was introduced and an optimized algorithm was proposed based on the faster RCNN （region-based convolutional neural network）^[1] framework for parameter tuning in this paper. The lightweight MobileNetv2 was utilized as the backbone network, and two additional area scales were added to the region proposal network（RPN）on top of the original anchor boxes. The ROI Align structure was employed in the detection part to reduce errors in feature mapping and pooling process. Experimental results showed that by using the faster RCNN object detection network pedestrian and vehicle detection tasks could effectively accomplished with overall good performance.