采用稀疏SIFT特征的车型识别方法
A Vehicle Classification Technique Based on Sparse Coding

针对实际应用中因图像清晰度低等因素导致的车型识别误差过大的问题,提出了一种基于稀疏尺度不变转换特征(sparse scale invariant feature transform,S？？SIFT)的车型识别方法。该方法用背景建模方法检测交通视频运动目标,提取目标SIFT特征;通过L1约束计算出SIFT特征的稀疏编码,并用最大池化方法降低稀疏编码维度,在线性SVM分类器中完成车型分类,弥补了背景建模方法识别误差过大、不具备车型分类功能的缺陷。经G36高速公路实际应用表明:算法对车辆场景识别率可达98％以上,车型识别准确率可达89％以上,对低清晰度、不同视角、雨雪、遮挡等场景有很好的鲁棒性;图像平均处理时间不超过40 ms,可满足系统对实时性的要求,在准确率和时间效率两方面均明显优于传统的SIFT方法和HOG方法。
A new method based on sparse scale invariant feature transform(S？？SIFT) is proposed to improve the vehicle recognition rate in environment such as low image quality. Moving objects are detected using a Gaussian mixture background subtraction model and SIFT features of the objects are calculated. Then, the sparse coding of SIFT features is obtained through L1 constraint. A max pooling strategy is introduced to reduce the dimension of the sparse coding. Finally, a linear support vector machine (SVM) is used to classify and to recognize the objects. The method solves the problems that the background modeling has a larger error rate and lacks function of vehicle classification. An application of the technique on G36 highway shows that the algorithm has an excellent result on different scenes such as low resolution, different camera angles, sleet and shade. The experimental results provide a more than 98% scene recognition rate, and a more than 89% classification accuracy rate. Moreover, the average time to process images is less than forty milliseconds, and it meets the real？？time requirement. It is concluded that the proposed method is better than the SIFT and the HOG methods on both accuracy and time efficiency