Crowd scene analysis is a highly focused research area of the intelligent surveillance. Due to severe occlusions between individuals, different illumination and diverse crowd distributions, it’s a challenging task in crowded environments. In this paper, two kinds of input feature maps of convolutional neural network (CNN) are proposed to characterize crowd motion for abnormal event detection in crowd videos. First, stacked optical flow map and streaklines mapare constructed based on optical flow field. Next, different kinds of feature maps are fed into CNN to train models which are then treated as a classifier for anomaly detection. Finally, the experiments, conducted on publicly available datasets, evaluate the results of different feature maps and show the effectiveness of the proposed method.