基于CNN-Transformer的自动泊车车位感知算法  

Algorithm for Parking Space Detection in Automatic Parking System Based on CNN-Transformer

作　　者：王玉龙 翁茂楠[1] 黄辉 覃小艺 Wang Yulong;Weng Maonan;Huang Hui;Qin Xiaoyi(Auto Engineering Research Institute,Guangzhou Automobile Group Co.,Ltd.,Guangzhou 510641;State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body,Hunan University,Changsha 410082)

机构地区：[1]广州汽车集团股份有限公司汽车工程研究院,广州510641 [2]湖南大学,汽车车身先进设计制造国家重点实验室,长沙410082

出　　处：《汽车技术》2024年第8期1-6,共6页Automobile Technology

基　　金：湖南大学汽车车身先进设计制造国家重点实验室开放基金项目(31825011)。

摘　　要：为提高自动泊车成功率及准确性,首先基于卷积神经网络(CNN)模型对输入图像进行特征提取,然后利用Transfomer模型的“编码-解码”机制对CNN提取到的图像特征平铺后进行计算推理,通过前馈神经网络得到目标预测结果,最后基于180°广角鱼眼图像进行推理识别,车位角中心点和空车位入口中心点均采用二维坐标表示,降低了输出信息的冗余,优化了模型结构。测试结果表明,该算法能够较好地适应不同车位线划线方式和不同的自然环境,目标感知的召回率达到98%,车位角中心点定位平均误差小于3 cm,满足泊车系统对车位感知的鲁棒性、实时性和准确性要求。In order to improve success rate and accuracy of automatic parking,firstly,the input image features were extracted based on Convolutional Neural Network(CNN)model,and then the encoding-decoding mechanism of Transfomer model was used to tile the image features extracted by CNN for calculation and inference.Finally,the target prediction results were obtained by feedforward neural network.In this paper,fisheye images were used to recognize the target.The center point of the parking angle and the center point of the empty parking entrance were expressed by two-dimensional coordinate points,which reduced the redundancy of the output information and optimized the model structure.The test results show that the algorithm can better adapt to different parking space line marking mode and different natural environment,with the recall rate of target perception reaches 98%,and the average error of parking space corner center location is less than 3 cm,which meets the requirements of real-time application for robustness,real-time and accuracy.

关 键 词：自动泊车 车位检测 视觉增强 卷积神经网络 TRANSFORMER 

分 类 号：U463.3[机械工程—车辆工程]