基于热图回归的停车位检测算法研究

Research on Parking Space Detection Algorithm Based on Heatmap Regression

实际停车场景存在环视视野受限或车位标记被遮挡等问题,现有方法难以有效地检测停车位。为提高自动驾驶车辆在复杂场景中停车位检测的可靠性,提出一种基于热图回归的停车位检测方法,将停车位简化为关键点、入口线和分隔线3个部分,首先将输入图像通过ResNet50和上采样模块提取关键点特征图;然后利用关键点估计模块生成关键点热图以及关键点坐标偏移、方向和长度属性,通过MaxPool提取热图中的峰值,输出置信度大于阈值的关键点;最后整合关键点的属性得到停车位,实现了端到端的停车位检测。实验表明:该方法在视野受限或被遮挡场景下具有良好的检测效果,在ps2.0测试集中实现99.01%的精度和98.27%的召回率,单帧检测速度为19.0 ms。

In actual parking scenarios,there are problems such as limited surrounding visibility or obscured parking space markings,and existing methods are difficult to effectively detect parking spaces.In order to improve the reliability of parking space detection for autonomous vehicles in complex scenes,a parking space detection method based on heatmap regression is proposed.The parking space is simplified into three parts:key points,entrance lines,and dividing lines.First,the input image is processed through ResNet50 and upsampling modules to extract key point feature maps.Then,the key point estimation module is used to generate key point heatmaps and key point coordinate offsets,directions,and length attributes.The peak values in the heatmaps are extracted using MaxPool,and key points with confidence greater than the threshold are output.Finally,the properties of the key points are integrated to obtain the parking space and achieve end-to-end parking space detection.Experiments show that this method has good detection performance in scenes with limited visibility or obscured areas,achieving a precision of 99.01%and a recall rate of 98.27%in the ps2.0 test set,with a single-frame detection speed of 19.0 ms.