【摘要】 针对深度模型在网络深层易丢失细节特征导致对小尺度目标检测效果差的问题,提出一种基于YOLOv3算法的小尺度交通信号灯检测模型。首先,采用跨越式特征融合提升浅层特征图的语义能力、减少过度融合产生的冗余信息;然后,采用K-means算法聚类出适合交通信号灯尺寸的新先验框,再采用线性缩放机制对新先验框离散以提升IoU。经过Bosch Small Traffic Lights Dataset测试集测试表明:所设计的新模型相较原YOLOv3模型,其mAP提升约9%,Green-AP提升5%,Red-AP提升30%,检测速度达24 fps,满足交通信号灯实时检测需求。另外,提出一种YOLOv3与OCR结合识别倒计时数字灯的方法,该方法在自制测试集上的识别精度达91%。更多还原

【Abstract】 Aiming at the problem that the depth model is easy to lose detail features in deep layer of network, resulting in poor detection effect for small-scale objects, propose a small-scale traffic light detection model based on the YOLOv3 algorithm.Firstly, leapfrog feature fusion method is used to improve the semantic ability of shallow layer feature maps and reduce the redundant information generated by excessive fusion.Then, K-means algorithm is used to cluster new priori box suitable for traffic light size, and then, a linear scaling mechanism is used to discretize the new priori box to improve IoU.Test by Bosch Small Traffic Lights Dataset test show that, compared with the original YOLOv3 model, the designed new model has an increase of 9 % in mAP,5 % in Green-AP,30 % in Red-AP,and the detection speed reach 24 fps, meeting the real-time detection requirements of traffic lights.In addition, a method combining YOLOv3 and OCR to recognition countdown digital lights is proposed, and recognition precision of this method on the self-made test set reaches 91 %.更多还原