基于TBGC的航拍视频车辆检测算法

Vehicle detection algorithm based on TBGC in aerial video

针对移动航拍视频中车辆检测准确度低的问题,提出一种基于三邻域点二值梯度轮廓(Three-neighbor-point Binary Gradient Contour,TBGC)特征的航拍车辆检测算法。对相邻帧图像进行SURF(speeded-up robust features)特征点提取匹配,利用角度判别剔除错误匹配点完成图像配准,采用帧间差分获得运动目标的候选区域。由于传统二值梯度轮廓(Binary Gradient Contours,BGC)特征忽略中心像素特性,提出基于3×3邻域相邻像素点量化操作的TBGC特征。提取候选区域的TBGC特征,并利用支持向量机(Support Vector Machine, SVM)完成最终的航拍视频车辆检测。实验中利用提出的TBGC特征在8个数据集上分别与BGC1、LBP、HOG特征进行对比实验,实验结果表明TBGC算法的检测率明显优于传统经典算法,平均检测率为93.09%,并且具有较好的鲁棒性。

To improve the vehicle detection precision in aerial video, a vehicle detection method for aerial video is proposed based on three-neighbor-point binary gradient contour(TBGC) features. First, the extraction of speed-up robust features(SURF) was extracted and matched for adjacent frame images, and error-matching points were eliminated by angle discrimination to realize image registration. Then, the candidate regions of the moving target were obtained according to the frame difference. Since the traditional binary gradient contour(BGC) features ignore the pixel characteristics at the center, the TBGC features based on the comparison of adjacent pixel points in a 3 × 3 neighborhood were proposed. In this way, the TBGC features in candidate regions were extracted, and support vector machine(SVM) was used to complete the vehicle detection in aerial video at last. In the experiment, the proposed TBGC features were compared with features of BGC1, LBP, and HOG on eight datasets respectively. Experimental results show that the detection rate of TBGC algorithm is obviously better than the traditional classical algorithms, with an average detection rate of 93.09% and stronger robustness.