基于全景环视图像的停车位检测算法

Parking slot detection based on surround-view image

为解决自主泊车过程中的停车位识别问题,该文基于全景环视图,提出了融合改进Transformer模块和卷积的混合编码-解码结构网络模型.使用上述模型对全景环视图中停车位上的“T”形和“L”形角点进行检测,并回归得到角点的位置、方向以及类型信息.然后把提取出的角点输入推理模块,根据角点对之间的种类和几何关系完成停车位检测并计算得到停车位位置.最后,使用ps2.0数据集在多个情景下进行了实验,停车位检测的精确率和召回率分别到达了99.27%和99.22%,角点位置误差RMSE在2 cm以下.对每张图片角点检测的计算量约2.71 GFLOP,对应的检测时间约40 ms.

To solve the problem regarding parking slot detection during automated valet parking(AVP)process,a hybrid encode-decode structured network model was presented using improved Transformer module and convolution.“T”and“L”shaped parking slot corner points were detected with above mentioned network and position,direction and type information of corner points were obtained by regression.Then,corner points were put into an inference module which could detect parking slot base on types and geometric relations between corner point pairs.Finally,multiple experiments regarding a variety of scenarios were performed on ps2.0 parking slot dataset,with precision rate and recall rate of detection reaching 99.36% and 99.31%,respectively,while RMSE of position error was only 0.87 pixel.The computational cost to detect corner points of a single image was around 2.71 GFLOPs and corresponding detection time was about 40 ms.