目标检测中语义约束检查算法的研究与实现

基本的目标检测任务是在图像中识别目标，并标注目标的类别和位置信息。但是，很多应用中的目标检测任务常常带有语义约束，典型的包括单类别目标的数量约束和多个目标之间的空间位置约束。如在基于视频的生产安全监控系统中，目标检测不仅要识别和标定安全防护装备，还要检测这些安全防护装备是否被规范穿戴。提出了一种目标检测中语义约束检查算法，定义一种语义约束的模型，然后对图像进行带有语义信息的目标检测，最终对目标检测结果与语义约束进行一致性判定。以电力施工防护装备检查的实际需求和现场安监视频为例，验证了所提出的目标检测中语义约束检查算法的有效性。

Research and Implementation of Semantic Constraint Verification Algorithm in Object Detection

The aim of basic object detection consists in identifying objects in an image, as well as classifying and localizing them. However, in many application scenarios, object detection is often semantically constrained. These constraints typically include the quantitative constraint to a single category object and spatial position constraint to multiple objects. For an instance, in a video surveillance system for manufacture safety, the object detection is designed to not only identify and label safety protective equipment, but also verify the appropriate usage of it under specifications. This paper proposes a checking algorithm with semantic constraint for object detection. Firstly, a semantic constraint model is defined, and then objects in an image are detected with semantic information. Finally, the consistency of the preliminary detecting result and the semantic constraint is verified. The validity of thissemantic constraint verification algorithm is evaluated by an application instance on protective equipment detection in electricity construction and on-site safety surveillance video.