Many studies on the object detection emphasizes the accuracy of the algorithms themselves, while the requirement of real-time processing can be addressed by the usage of “you only look once” (YOLO) model. However, the reliably of machine vision is still a problem since some practical issues are not addressed properly, such as variation of light intensity, reflection of light on the surface and interference of shooting background. In this paper, we address above problems by developing a vision system with YOLO algorithm for object detection, segmentation and localization. A segmentation approach is adopted on the model outputs to extract the object to be detected from the background, under the premise of enhancing the adaptability of the YOLO model to environmental changes. Thus, the influence of background and light-sensitive factors on localization is removed even in extreme lighting conditions. An experimental platform is built based on a pair of low-cost cameras, which verifies the effectiveness of proposed method.

许多关于目标检测的研究都强调算法本身的准确性，而实时处理的要求可以通过使用“you only look once”（YOLO）模型来解决。然而，机器视觉的可靠性仍然是一个问题，因为一些实际问题没有得到妥善解决，例如光强度的变化，光在表面的反射和拍摄背景的干扰。在本文中，我们通过使用 YOLO 算法开发用于目标检测、分割和定位的视觉系统来解决上述问题。在增强YOLO模型对环境变化的适应性的前提下，对模型输出采用分割的方法从背景中提取待检测对象。因此，即使在极端照明条件下，背景和光敏因素对定位的影响也被消除了。基于一对低成本相机搭建了实验平台，验证了所提方法的有效性。