基于YOLOv4算法在车辆检测中的应用

Application of YOLOv4 Algorithm in Vehicle Detection

为解决车辆识别中由于拍摄角度和距离的不同,导致成像后的车辆尺寸较小和车辆存在不同程度的遮挡,从而产生车辆的错检和漏检等问题,在单阶段目标检测网络YOLOv4(You Only Look Once version 4)算法的基础上,提出了基于注意力机制的递归YOLOv4目标检测算法,即RC-YOLOv4(Recursive and CBAM You Only Look Once version 4)算法。为提高算法对成像后小尺寸车辆的检测能力,在YOLOv4算法加入CBAM(Convolutional Block Attention Module)模块,该模块结合了通道和空间注意力机制,能帮助网络模型更加关注检测图像中的重点信息和小目标信息。针对车辆部分遮挡的检测问题,采用递归特征金字塔(RFP:Recursive Feature Pyramid)结构加强模型对深层特征信息提取能力,RFP结构类似于选择性增强或抑制神经元激活的人类视觉感知,将主干网络提取到的特征递归融合,然后反馈给主干网络,多次特征融合增强网络对上下文语义信息的提取整合能力。提高了对遮挡车辆的检测精度。实验结果表明,在自制车辆检测数据集上,RC-YOLOv4算法相比于YOLOv4在平均精度均指标上提高了12.69%,同时检测速度也能满足实时性要求。

In vehicle recognition, due to different shooting angles and distances, the size of the imaged vehicle is smaller and the vehicle has different degrees of occlusion, resulting in detection error and missed detection. In order to solve this problem, based on the single stage target detection network YOLOv4(You Only Look Once version 4) algorithm, a recursive YOLOv4 target detection algorithm is proposed based on attention mechanism, namely RC-YOLOv4 algorithm. In order to improve the detection capability of the algorithm for small size vehicles after imaging, the CBAM(Convolutional Block Attention Module) module is added to YOLOv4 algorithm. This module combines the channel and spatial attention mechanism, which can help the network model pay more attention to the key information and small target information in the detected image. For the detection of partial occlusion of vehicles, a RFP(Recursive Feature Pyramid) structure is adopted to enhance the model's ability to extract deep feature information. The RFP structure is similar to the human visual perception that selectively enhances or inhibits the activation of neurons. The features extracted from the backbone network are recursively fused and then fed back to the backbone network. Multiple feature fusion improves the network's ability to extract and integrate contextual semantic information. It improves the detection accuracy of occluded vehicles. The experimental results show that the average precision of RC-YOLOv4(Recursive and CBAM You Only Look Once version 4) algorithm is 12.69% higher than YOLOv4 algorithm on the self-made vehicle detection data set, and the detection speed can also meet the real-time requirements.