改进的RetinaNet模型的车辆目标检测

摘要

目前,在智能交通领域使用深度学习方法进行车辆目标检测已成为研究热点.针对传统机器学习方法的性能易受光照、角度、图像质量等外界因素影响,检测步骤繁琐等问题,通过对当下经典的一阶目标检测模型和二阶目标检测模型进行分析,提出了一种基于改进的一阶目标检测模型RetinaNet的车辆目标检测方法,使用深度残差网络自主获取图像特征,融合MobileNet网络结构进行模型加速,把复杂交通场景下的目标检测问题转化为车辆类型的三分类问题,利用KITTI数据集进行训练,并使用实际场景中的图像进行测试.实验结果表明,改进的RetinaNet模型在保证检测时间的情况下,相比原RetinaNet模型MAP值提高了2.2个百分点.%At present, in the field of intelligent transportation, vehicle target detection has become a research hotspot by using deep learning method. As traditional machine learning method is easily affected by external factors such as illumi-nation, angle and image quality, and the cumbersome detection steps, current one-stage target detection models and two-stage target detection models are analyzed, and vehicle target detection method is proposed by on one-stage target detection model and is named RetinaNet, which uses deep residual network to acquire image features autonomously, integrates MobileNet network structure to accelerate the model and transforms the target detection problem in complex traffic scenarios into the three-category problem of vehicle type. KITTI data set is adopted to train and actual scenes images are applied to test. Experimental results show that the proposed method improves the MAP value by 2.2 percentage points in comparison with original RetinaNet model.