基于扩张卷积特征自适应融合的复杂驾驶场景目标检测

Object detection in complex driving scene based on dilated convolution feature adaptive fusion

针对复杂驾驶场景下的目标检测问题,提出一种基于扩张卷积特征自适应融合的目标检测算法.采用单阶段目标检测网络RetinaNet作为基本框架,其包含卷积特征提取、多尺度特征融合以及目标分类与回归子网.为提高网络对多尺度特征的提取能力,设计了基于不同扩张率组合的残差卷积分支模块,以获取不同感受野下的目标特征图;然后,将不同尺度下的特征通过网络自适应学习的参数融合后输出,用于后续的目标预测;最后在大规模且多样化的复杂驾驶场景数据集BDD100K上进行实验.结果表明,利用扩张残差卷积分支模块与特征自适应融合算法能够分别将网络的平均精度均值由0.330提升至0.338与0.344,并在采用扩张卷积特征自适应融合的情况下达到了0.349.所提算法能够有效提升目标检测算法在复杂驾驶场景下的检测性能.

Aiming at the problem of object detection in complex driving scene,an algorithm based on dilated convolution feature adaptive fusion was proposed.The single-stage object detection network RetinaNet was adopted as the basic framework,including feature extraction,multi-scale feature fusion,and object classification and regression subnets.To improve the capability of multi-scale feature extraction,a residual convolution branch module based on different expansion rate combinations was designed to obtain target feature maps in different receptive fields.Then,features at different scales were adaptively fused through the parameters learned by network itself,and the output fusion feature maps were used for subsequent predictions.Finally,experiments were carried out on the large-scale and diversified complex driving scene dataset BDD100K.Experimental results show that using the dilated residual convolution branch module and the feature adaptive fusion algorithm,the mean average precision of the network is increased from 0.330 to 0.338 and 0.344,respectively.And it reaches 0.349 with the dilated convolution feature adaptive fusion.The algorithm can effectively improve the performance of the object detection algorithm in complex driving scene.