Abstract: Vehicle detection is the premise of the automotive anti-collision warning.This paper presents a multi-feature-combined approach to improve the robustness of the vehicle detection in real-time.The approach dose not depend on the lane detection for it is based on the grads feature of the shadow which shows the candidate vehicle regions and it eliminates the noises of the corresponding area by the method of differential box counting.Then，the accurate vehicle area can be located by analyzing the information of vehicle’s horizontal edge feature in the candidate vehicle region.Finally，Kalman filters are used to track the candidate vehicle which will be validated by normalized-mutual-information feature.The result of the experiment has shown that the method provides a robust approach，which can effectively detect the front vehicles in complex traffic circumstances in real time.

Key words: vehicle detection, grads, differential box counting, edge detection, Kalman filter

摘要： 车辆检测是汽车防碰撞预警的前提，为了提高前方车辆检测的实时性和鲁棒性，提出一种结合多特征的前方车辆检测跟踪方法。该方法不依赖车道检测，利用车底部阴影的梯度特征确定可能存在车辆的区域，使用差分盒子维计算对应区域的分形维数来排除噪声，根据车辆的水平边缘特征信息精确定位，通过卡尔曼滤波器跟踪检测到的目标，利用归一化转动惯量做车辆验证。实验结果表明，该方法能够在多种交通环境中实时有效地检测前方车辆。

关键词: 车辆检测, 梯度, 差分盒子维, 边缘检测, 卡尔曼滤波