中文摘要：

基于背景建模的交通流参数视频检测方法易受车辆遮挡、光线变化、雨雪等外界环境条件干扰等而误判;基于机器学习方法的机理不同,改变帧间像素动态变化解析的方式,而着重于车辆样本空间的目标识别提取,具有抗干扰优势.为此,在深度学习的基础上,提出了基于稀疏特征的交通流检测算法.其构建高斯混合背景模型,提取交通视频的运动目标,以稀疏编码处理目标的尺度不变特征来获得稀疏特征;经最大池化的稀疏特征维度降低、线性支持向量机训练、背景建模方法误判样本去除,从而计算获得交通流参数.测试结果表明：与背景建模方法比较,该算法去除了60%~80%误判,准确率提高了12%~17%,对外界干扰大、分辨率不高的视频图像优势显著.

英文摘要：

The accuracy of traditional traffic flow detection based on background modeling is not high,because it is susceptible to external environmental interference.In this paper,a video-based traffic flow detection algorithm using sparse feature is propsed.Moving objects are firstly detected with Gaussian mixture background modeling method.Then the scale-invariant feature transform（SIFT）of the objects are calculated and sparse coding is used to handle the SIFT.For the purpose of reducing sparse feature dimensions,the max pooling strategy is introduced.Afterwards,the pooling-process results are used to train the linear support vector machine（SVM）.Finally,we use the SVM to remove the negative samples,thus calculate better traffic flow parameters.Test results show that our algorithm has outstanding performance on low resolution images：the negative sample removal rate is 60% to 80%,accuracy improved 12% to 17% higher than background modeling method.