Identification of abnormal behaviors in surveillance videos of crowds plays an important role in public security monitoring. However, detecting abnormal crowd behaviors is challenging in that movements of individuals are usually random and unpredictable, and the occlusions caused by over-crowding make the task more difficult. In this paper, we introduce thermodynamic micro-statistics theory to detect and localize abnormal behaviors in crowded scenes based on Boltzmann Entropy. For this purpose, the scene of interest is modeled as moving particles turned out from a general optical flow algorithm. The particles are grouped into a set of prototypes according to their speeds and directions of moving, and a histogram is established to figure out how the particles distribute over the prototypes. Here, Boltzmann Entropy is computed from the histogram for each video clip to characterize the chaos degree of crowd motion. By means of such feature extraction, the crowd motion patterns can be represented as a time series. We find that when most people behave anomaly in an area under surveillance, the corresponding entropy value will increase remarkably in comparison with those of normal cases. This motives us to make use of Boltzmann Entropy to distinguish the collective behaviors of people under emergent circumstances from their normal behaviors by evaluating how significantly the current feature value fits into the Gaussian model of normal cases. We validate our method extensively for anomaly detection and localization. The experimental results show promising performance compared with the state of the art methods.