基于改进的几何约束算法与卷积神经网络的车辆检测

Vehicle Detection Based on Improved Geometric Constraint Algorithm and Convolution Neural Network

论文提出了一种基于改进的几何约束算法有效结合卷积神经网络(CNN)的车辆检测方法。首先对几何约束算法进行改进,避免该算法重复的矩阵运算,从而进一步提高该算法效率。根据改进的几何约束算法计算出车辆的感兴趣区域,然后在该区域内提取Haar-like特征,通过Adaboost分类器初步检测得到候选框。之后用训练好的卷积神经网络模型对目标候选框进行分类。实验结果表明,该方法能够有效地减少车辆检测时间,提高车辆检测的精度,并且对多种光照条件,部分遮挡,姿态变化等具有一定的鲁棒性。

In this paper,a vehicle detection method based on improved geometric constraint algorithm combined with convolu?tion neural network is proposed.Firstly,the geometric constraint algorithm is improved to avoid the matrix operation,in order to fur?ther improve the efficiency of the algorithm.According to the improved geometric constraint algorithm,the region of interest of the vehicle is obtained.And the Haar-like feature is extracted in the region,then the Adaboost classifier is used for the preliminary de?tection.The convolution neural network model is trained by the vehicle training set.The convolution neural network model is used to classify candidate regions.The experimental results show that this method can reduce the vehicle detection time and improve the ac?curacy of vehicle detection effectively,and it is robust for a variety of lighting conditions,partial occlusion and posture.