基于自适应融合的实时车辆检测

Real-time Vehicle Detection Based on Adaptive Fusion

针对传统的车辆检测技术检测速度慢和精度低的问题,提出了一种融合注意力的自适应金字塔网络的交通目标检测算法(fusion attentiont adaptive pyramid network,FAAP-Net),可以显著降低交通事故的发生率。为了降低计算复杂度,设计了一种轻量级的互补池化结构(CPS),该结构在宽度和高度上采用了两组不同的池化组合,在保持高精度的同时,显著降低了网络的浮点运算数(GFLOPs)和参数量。为了解决智能交通系统特征图生成过程中的信息损失问题,通过将自适应注意力模块(AAM)和特征增强模块(FEM)引入自适应融合特征金字塔网络(AF-FPN),以融入车辆检测的形状特征。针对车辆细节特征表征弱的问题,引入了一种按通道维度分组的注意力(SA)机制,以增强主干网络对不同车辆检测细节特征的关注,有效提取车辆细节的显著特征。在BDD100K数据集上的实验结果表明,FAAP-Net算法相比于传统算法,平均精度从30.3%提升到43.7%。

A traffic target detection algorithm,fusion attention adaptive pyramid network(FAAP-Net),is proposed to address the issues of slow speed and low accuracy in traditional vehicle detection techniques,significantly reducing the occurrence of traffic accidents.To mitigate computational complexity,a lightweight complementary pooling structure(CPS)is designed,employing two sets of different pooling combinations in width and height,which maintains a high precision while significantly reducing the floating point operations per second(GFLOPs)and the parameter count of the network.Addressing the information loss during intelligent traffic system feature map generation,the adaptive fusion feature pyramid network(AF-FPN)incorporates the adaptive attention module(AAM)and the feature enhancement module(FEM)to integrate shape features for vehicle detection.Lastly,to address the weak representation of vehicle detail features,a channelwise grouped attention(SA)mechanism is introduced,enhancing the focus of the backbone network on various vehicle detection details and effectively extracting significant features.The experimental results on the BDD100K dataset demonstrate that the FAAP-Net algorithm achieves a notable improvement,increasing the average precision from 30.3%to 43.7%.