结合注意力机制的车型检测算法

Vehicle Detection Algorithm Combined with Attention Mechanism

针对目标检测算法应用在车辆类型检测的场景中,检测速度较快,但检测精度相对较低的问题,该文对CenterNet算法进行改进。首先,使用ResNet作为主干网对车型图像进行特征提取,并在特征提取网络中引入通道注意力和空间注意力,对不同通道以及不同位置的特征进行权重划分,获取更多需要关注的特征,抑制无用的特征,进而提升车型检测算法的分类及定位准确率;其次,针对小目标车型检测精度不高的问题,将不同尺度车型特征进行融合,更好地提取细粒度车型特征,提升检测精度。为验证结合注意力机制的车型检测算法的有效性,在KITTI车型数据集和BIT-Vehicle数据集上进行实验,mAP值分别达到94.6%和95.5%。结果表明改进后的算法模型在检测速度影响较小的情况下检测精度得到显著提升。

Aiming at the problem of fast detection speed and relatively low detection accuracy for the object detection algorithm in the scene of vehicle type detection,we improve the CenterNet algorithm.Firstly,ResNet is used as the backbone network to perform feature extraction on vehicle images,and the channel attention and spatial attention are introduced into the feature extraction network to carry out the weight division of the features in different channels and at different positions,to obtain more features that need attention and suppress useless features,thus improving the classification and positioning accuracy of vehicle detection algorithms.Secondly,in view of the problem of low detection accuracy of small target vehicles,we integrate the features of different scale vehicle to better extract fine-grained vehicle features and improve detection accuracy.To verify the effectiveness of the vehicle detection algorithm combined with the attention mechanism,experiments were conducted on the dataset of KITTI and BIT-Vehicle.The mAP values reached 94.6%and 95.5%respectively.The results show that the improved algorithm can significantly improve the detection accuracy with little influence on the detection speed.