基于学习算法SSD的实时道路拥堵检测

Real-time Road Congestion Detection Based on Learning Algorithm SSD

随着人们生活水平的不断提高,道路车辆拥堵情况愈发严重。如何实时、精确地检测出道路车辆,对于解决道路拥堵问题具有重要意义,GPU和人工智能技术的飞速发展为其提供了可靠的解决方案。研究分析传统目标识别算法、基于候选区域深度学习的目标提取算法RCNN和基于回归的深度学习目标检测YOLO,最终确定采用基于卷积神经网络的实时目标检测算法SSD。首先调用VGG16网络模型在ILSVRC CLS-LOC数据库上预训练生成初始网络模型,进而设置超参数并在自身数据集上进行再训练,生成新的网络检测模型,然后将训练和测试部署到深度学习框架Caffe上加以实现。通过在数据库COCO、VOC2012上的测试结果表明,该模型检测精度为76%左右,处理速度为26FPS。同时通过道路路口的实地车辆检测,显示该算法能够实时、精确地检测出道路车辆,为道路拥堵情况判定提供可靠数据。

With the continuous improvement of the people′s living standard,the congestion of road vehicles becomes more and more serious.How to accurately detect road vehicles in real time is of great significance to solve the problem of road congestion.The rapid development of GPU and artificial intelligence provides a reliable solution.We analyse the traditional target recognition algorithm,target extraction algorithm RCNN based on the candidate region deep learning and regression-based deep learning target detection YOLO,and finally we adopt the real-time target detection algorithm SSD based on convolutional neural network.First we call the VGG16 network model to pre-train the initial network model on the ILSVRC CLS-LOC database to set up hyper-parameters and retrain on their own data sets to generate a new network detection model and then deploy training and testing to the deep learning framework Caffe.Tests on the database COCO,VOC2012 showed that the model detection accuracy reached about 76%,the processing speed reached 26 FPS.the final road junctions The field of vehicle detection in the road junctions proves that the algorithm can realise real-time accurate detection of road vehicles so as to provide reliable data about road congestion.