为满足智能农业车辆的精确导航,提出基于立体视觉的车辆实时运动检测方案.该方案通过多线程特征点检测提高传统SIFT特征检测算法的效率,通过归一化综合距离法剔除误匹配的特征点,最后通过相邻时刻同一特征点坐标的变化反推车辆的运动.试验表明:多线程SIFT特征点检测能够缩短检测时间,提高计算效率.归一化综合距离法能够有效剔除传统SIFT算法的误匹配点.当车速为0.8 m/s,图像采集频率为5 Hz时,车辆在x方向和z方向单次测量误差小于0.004 5 m,当持续运动时间达到10s时,2个方向累积测量误差均小于0.15m.%A real-time motion detection method based on stereo vision was designed in order to meet the need of precise navigation for intelligent agricultural vehicle.Multi-thread feature points detection was used to improve the efficiency of traditional SIFT algorithm.Normalized comprehensive distance algorithm (NCDA) was used to delete the error matched points.Finally,vehicle motion was calculated by the position variation of feature points in adjacent moment.Experiments showed that the average detection time spent was reduced and computing efficiency was raised up.Error matched points were successfully recognized and erased by NCDA.Measurement errors in one time were less than 0.004 5 m in x-and z-axis when the speed was 0.8 m/s and image grabbing frequency was 5 Hz.Measurement errors in x-and z-axis were less than 0.15 m when the vehicle kept moving for 10 s.
展开▼
