基于深度学习的停车位检测研究

Vehicle Parking Slot Detection Research based on Deep Learning

随着智能网联汽车的发展,越来越多的学者投身于L4级以上的稳定的自动驾驶算法研究中来。自动泊车系统作为智能网联汽车的一项重要功能,能够在有效提升驾驶体验的同时,降低由于复杂地段的泊车困难带来的交通事故和经济损失,因此自动泊车在学术界和工业界掀起了研究热潮。传统的自动泊车系统中对于车位的感知依赖于超声波雷达,并且对车位空间结构有诸多限制。由于复杂的视觉环境和环视图像上停车位的不完整显示,基于视觉的停车位检测是一项重大挑战。本文提出了一种基于卷积神经网络(CNN)的车位检测算法,设计适用于车载环视图像的多重沙漏网络,并引入一种策略选择最佳感受野,从而联合检测停车位的角和线特征。所提出的方法达到了较高的精度和召回率,在搭载GPU的嵌入式移动终端可以达到30 FPS的实时性和较高的精准度。

With the development of intelligent connected vehicle(ICV),an increasing number of scholars are devoted to the research of stable autonomous driving algorithms above L4 level.As a significant component of ICV,automatic parking system can effectively improve the driving experience while reducing traffic accidents and economic losses caused by parking difficulties in complex areas,which has set off a research boom in academia and industry.Traditional automatic parking system relies on ultrasonic radar,and there are many restrictions on the space structure of the parking slot.However,computer vision based parking slot detection is also a difficult challenge due to the complex environment and the block-out of parking slot in bird eye view(BEV) images.In this paper,a parking slot detection algorithm based on convolution neural network(CNN) is proposed which is composed of multiple hourglass network(MHN).A specific strategy is introduced to select the best convolution kernel candidates,so that the model can extract joint-cross-side features of parking slots.The proposed method achieves high precision and recall,and achieves 30FPS in embedded mobile terminals equipped with GPU.