基于改进YOLOv5s的面向自动驾驶场景的道路目标检测算法

Road target detection algorithm for autonomous driving scenarios based on improved YOLOv5s

在复杂道路场景中检测车辆、行人、自行车等目标时,存在因多尺度目标及部分遮挡易造成漏检及误检等情况,提出一种基于改进YOLOv5s的面向自动驾驶场景的道路目标检测算法。首先,利用深度可分离卷积替换部分普通卷积,减少模型的参数量以提升检测速度。其次,在特征融合网络中引入基于感受野模块(receptive field block,RFB)改进的RFB-s,通过模仿人类视觉感知,增强特征图的有效感受野区域,提高网络特征表达能力及对目标特征的可辨识性。最后,使用自适应空间特征融合(adaptively spatial feature fusion,ASFF)方式以提升PANet对多尺度特征融合的效果。实验结果表明,在PASCAL VOC数据集上,所提算法检测平均精度均值相较于YOLOv5s提高1.71个百分点,达到84.01%,在满足自动驾驶汽车实时性要求的前提下,在一定程度上减少目标检测时的误检及漏检情况,有效提升模型在复杂驾驶场景下的检测性能。

When vehicles,pedestrians,bicycles,and other targets are detected in complex road scenes,the existence of multiscale targets and partial occlusions may easily cause missed and false detections.In this paper,a road target detection algorithm is proposed based on improved YOLOv5s,orienting to autonomous driving scenarios.First,depthwise separable convolution is used to replace partial ordinary convolutions to reduce the number of parameters of the model to improve the detection speed.An improved RFB-s based on receptive field block(RFB)is introduced into the feature fusion network to enhance the effective receptive field area of the feature map,improving the network feature expression capability and the recognizability of the target features by imitating human visual perception.Finally,an adaptive spatial feature fusion method is used to enhance the effect of PANet on multiscale feature fusion.The experimental results reveal that,on the PASCAL VOC dataset,compared with YOLOv5s,the mean value of the average detection precision of the proposed algorithm is improved by 1.71%,reaching 84.01%.Under the premise of meeting the real-time requirement of autonomous driving vehicles,this algorithm has reduced false and missed detections in the target detection to a certain extent,effectively improving the detection performance of the model in complex driving scenarios.