基于优化YOLO算法的多约束条件下视觉伺服系统目标的精准定位

Accurate Target Location of Visual Servo System Under Multiple Constraints Based on Optimized YOLO Algorithm

在实际工业环境中完成系统的基本硬件部署后,对在复杂的多约束条件下具有目标特征的图像进行有效识别是视觉伺服系统面临的关键问题。因此,结合视觉伺服系统在复杂环境中目标检测的特点,利用基于YOLO训练方法的目标检测和特征提取一体化网络结构,提出一种将上一帧目标位置与卡尔曼滤波得到目标在当前帧的预测位置做关键点匹配,从而判断预测位置是否存在目标的方法。该方法能够有效提高图像目标特征的检测效率,目的在于克服现有技术在检测速度和成功率上的不足,设计服务于机器人视觉伺服系统总体任务目标的检测框架。对所提方法进行了验证,实验表明:在模拟常规干扰项约束下,利用所提方法目标图像检出成功率可以提升5%以上。

After completing the basic hardware deployment of the system in actual industrial environment,effective recognition of images with target characteristics under complex multi-constraint conditions is the key problem that visual servo system facing.Therefore,combined with characteristics of target detection in complex environment of visual servo system,and with integrated network structure of target detection and feature extraction based on YOLO training,a method is proposed to match target position in previous frame with the predicted position of the target in current frame obtained by Kalman filter as key point,in order to judge whether there is the target in predicted position.By using the method,the detection efficiency of image target features can be effectively improved,aiming to overcome shortcomings of existing technologies in detection speed and success rate,and to design a detection framework serving the overall task target of robot visual servo system.The proposed method is verified,and the experiment shows that the success rate of target image detection can be improved by more than 5%using the proposed method under the constraint of simulating conventional interference.