In recent years, video based human abnormal behavior recognition algorithms have achieved certain research results. However, due to the large amount of data stored in surveillance videos and the long time span of videos, existing recognition methods are not suitable for detecting and recognizing abnormal actions in long videos or multiple pedestrians. To this end, a human abnormal behavior recognition model based on keyframe localization is proposed. Firstly, a keyframe localization network based on standardized flow and attention enhanced spatial temporal graph convolution is used to learn the probability distribution of normal frames, filter and extract sequences of abnormal frames (keyframes) in long videos, and use them as inputs for subsequent network models. Then, in order to better capture the motion characteristics and abnormal situations of human posture, a spatial temporal graph convolutional abnormal behavior recognition algorithm that integrates attention and enhanced residuals is proposed. The keyframe sequence is input into the model network to achieve efficient and accurate recognition of human abnormal behavior in surveillance videos. Validate the effectiveness of this method using publicly available and self built datasets. The experimental results show that the TOP-1 accuracy of human abnormal behavior recognition on the publicly available dataset ShanghaiTech Campus reaches 82.86%, and the TOP-5 accuracy reaches 98.10%. This method can better complete the recognition of human abnormal behavior in surveillance videos.