基于深度学习的驾驶场景关键目标检测与提取

Deep Learning Based Key Object Detection and Extraction for Driving Scene

包含目标识别与边界框选定的目标检测是无人驾驶视觉感知中的关键技术之一。采用基于深度计算机视觉组网络(VGGNet)的新型单次多框检测算法(SSD)进行驾驶环境中的关键目标检测、语义标注和目标框选;同时,针对具体驾驶场景,提出了改进的SSD_ARS算法。通过优化梯度更新算法、学习率下降策略和先验框生成策略,在提高平均检测精度的同时使得小目标类别的检测精度得到明显提升。在实际驾驶场景中9类关键目标的检测实验上验证了本文算法的有效性,实验结果表明,检测速度满足实时检测需求。

Object detection with object recognition and bounding box generation plays an important role in the visual perception of autonomous driving. With the rapid development of deep learning, a series of object detection algorithms based on deep convolution networks have been proposed. These vision-based algorithms have been widely used to deal with driving scenes or other complicated situations. In this paper, a novel detection algorithm, single shot multi-box detector(SSD)based on deep convolution network-visual geometry group network(VGG), is proposed to achieve target recognition, semantic annotation, and bounding box selection in driving scene. Moreover, an improved algorithm, termed as single shot multi-box detector with aspect ratio selection strategy(SSD_ARS), is also proposed for object detection in auto-driving. Momentum optimization is used in the gradient descent algorithm and the learning rate reduction strategy is optimized. Furthermore, the aspect ratio selection strategy is improved when generating the priori box. It is shown from the results on different selection strategies of aspect ratio when generating the priori box that the proposed algorithm in this work can effectively improve the accuracy of object detection, especially, the detection accuracy of small objects. Finally, the detection experiments of 9 significant object classes in the real driving scene are made, from which it is seen that the proposed algorithm performs well at near and medium distances. Besides, it is also demonstrated via the experiments on driving video that the detection speed of the proposed algorithm in this work can meet the requirements of real-time detection.