基于深度学习的公路货车车型识别

为判断公路货车车型，并提升货车车型识别的速度与精度，提出基于深度学习的方法对公路货车及其轮轴进行精细化目标检测；采用道路监控拍摄和网络爬取的方式获得了16 403张公路货车侧方图像，建立了货车侧方图像数据集，并采用Retinex理论和加入限制对比度的自适应直方图均衡化(CLAHE)等视觉增强方法预处理所采集图像中的光照不均图像和夜视图像；通过理论分析和对比试验选取单阶段检测网络YOLOv3作为公路货车车型识别的目标检测网络，并从调整先验框和模型输入大小以及引入注意力机制3个方面优化了检测模型；针对单帧图像可能同时出现多辆货车的情况，采用基于目标位置信息挖掘的算法分析了货车与轮轴的位置信息，提出一种通过轮轴中心点与货车预测框位置信息判定公路货车与轮轴隶属关系的方法。研究结果表明：图像经过预处理可显著增强车辆的特征信息，优化后检测模型的网络性能得到提高，通过对目标位置信息的挖掘与利用可以很好地解决货车车型判定问题；优化后的检测模型实时检测速度可达47帧·s^-1,对公路货车车型的识别综合准确率达到了94.4%。该方法实现了对公路货车车型的无接触、快速和准确识别，为公路货...

Type recognition of highway trucks based on deep learning

A deep learning-based fine target detection method for highway trucks and their wheel axles was proposed to determine the vehicle types of highway trucks and improve the recognization speed and accuracy of vehicle types of trucks. 16 403 side images of trucks were obtained by the road monitoring and network crawling to build a dataset of side images of trucks. The Retinex theory and visual enhancement methods, such as the contrast limited adaptive histogram equalization (CLAHE), were used to preprocess the uneven light images and night vision ones in the collected images. Theoretical analysis and comparative experiments were conducted. The one-stage detection network YOLOv3 was selected as the target detection network for the vehicle type recognition of highway trucks. Then, the detection model was optimized from three aspects, such as adjusting the sizes of prior box and model input and introducing the attention mechanism. For the case that multiple trucks might appear at the same time in a single image, an algorithm based on the mining of target location information was employed to analyze the truck and wheel axle location information. A method was proposed to determine the subordinate relationships between the highway truck and the wheel axle according to the location information of axle center points and truck prediction boxes. Research results show that the vehicle feature information can be enhanced significantly by the image preprocessing. The network performance of the detection model improves after the optimization. The issue of determining the vehicle types of trucks can be well solved by mining and leveraging the target location information. A real-time detection speed of the optimized detection model reaches 47 frames per second. The comprehensive accuracy in recognizing the vehicle types of highway trucks is 94.4%. The method realizes the non-contact, fast, and accurate recognition of vehicle types of highway trucks, provides a new means for the vehicle type recognition of highway trucks. Meeting the construction needs of intelligent traffic systems, the proposed method can be applied to further raise the road service level.