融合卷积神经网络和循环神经网络的车轮目标检测

Wheel detection integrating convolutional neural network and recurrent neural network

目标检测是基于视觉的目标定位关键技术。针对现有车轮检测方法对环境敏感问题,本文提出一种并联式融合循环神经网络和Faster R-CNN的车轮检测模型FusionRNN,借助RNN能够处理时序和CNN能够提取空间域隐性特征的优点,可提高实时性,减少参数量,使模型表达能力更强,同时具备分析序列化向量间语义关系和识别车轮几何特征的能力。该模型能在由激光雷达扫描得到的车轮三维点云投影图中准确检测出车轮位置,为基于AGV自动停车系统搬运车辆提供准确稳定的车辆位置信息。

Target detection is a key technology of target location based on vision.Aiming at the environmental sensitivity of the existing wheel detection methods,this paper proposes a wheel detection model integrating RNN and Faster R-CNN named FusionRNN.Leveraging RNN to deal with temporal information and CNN to extract the recessive characteristics in the spatial domain,it can improve the real-time performance,reduce the number of parameters,and make the model more expressive.At the same time,it has the ability to analyze the semantic relationship between serialization vectors and recognize the geometric features of wheels.The model can accurately detect the wheel in the three-dimensional point cloud scanned by the Li DAR,provide the vehicle position information for the intelligent manipulator used to grab the vehicle to the designated parking space,accurately obtain the vehicle’s center of gravity,and realize the safety,rapid capture and storage of the vehicle.