The invention discloses an abnormal behavior detection method based on time-space Laplacian Eigenmaps learning, belongs to the technical field of digital image processing, and relates to theoretical knowledge related to computer vision, mode identification, machine learning and data mining. An optical flow histogram is used to extract optical flow features from two adjacent frames of pictures, movement characteristic information in a monitoring scene is obtained, the movement characteristic information is clustered in a spectral clustering manner by using a video expression form of low-dimension space, the clustering amount and characteristic sets in different classifications are obtained, Hausdorff distance is applied to the characteristic sets to measure the similarity between the sets, a characteristic set which is different from those of other classifications obviously is searched, and thus, an abnormal behavior is detected. According to the invention, data in high-dimension space is re-expressed in a low-dimension space, the operational complexity is reduced, and abnormal behavior detection in a crowded scene is helped. The detection rate of abnormal behaviors is 73.52-78.45%, the omission rate 17.05-21.45%, and the false detection rate 4.5-6.1%.