Recently, using machine learning technology to realize abnormal behavior recognition in video surveillance to replace human monitoring has become a hot academic topic. In that case, constructing an efficient and unified framework for multi-type abnormal behavior recognition is a worthy topic in machine learning research. This research aims to design a lightweight recognition framework that can recognize various abnormal behaviors in real-time. We propose a Novel framewOrk for the Multi-type Abnormal BEhavior Recognition (NOMABER), which consists of three parts. Firstly, the improved image pre-processing module annotates the abnormal behaviors of image data sets. Secondly, the improved YOLOv5 module is used to identify the multi-type abnormal behaviors, and then the abnormal behaviors are classified by the output module. Finally, experiments on real data sets show that NOMABER is superior to the current methods of real-time performance, identification accuracy, and types of abnormal behaviors.