基于中心点检测的车流统计方法研究

Research on Traffic Flow Statistics Method Based on Center Point Detection

由于尺度变化、角度变化及遮挡问题,对图像中的车流数量进行估计具有较大的挑战。随着深度学习的发展,利用基于多列或多网络的神经网络模型来提取尺度相关的特征,以提升密度估计的精度,但是,这些模型在进行优化训练时较为复杂,且需要消耗巨大的计算资源。鉴于此,论文提出一种通过基于中心点检测的卷积操作,来提取车辆相关的特征信息,根据检测到的结果和标注数据构建Focal Loss,从而实现对高密度车流的估计。实验表明,该模型具有较高的精度和较好的鲁棒性。

Estimating the number of traffic in an image is a major challenge due to scale changes,angle changes,and occlusion problems. With the development of deep learning,neural network models based on multi-column or multi-network are used to extract scale-related features to improve the accuracy of density estimation. However,these models are more complex to perform optimization training and require huge computational resources. In view of this,this paper proposes a convolution operation based on center point detection to extract vehicle-related feature information,and constructs Focal Loss based on the detected results and the annotated data,thereby realizing the estimation of high-density traffic flow. Experiments show that the proposed model has higher precision and better robustness.