Abstract: Targeting the problems of dynamic background and diverse types of track abnormal objects in UAV (unmanned aerial vehicle) aorial vedio detection, an aerial vedio detection method of track abnormal objects based on CNN (convolutional neural network) is proposed. The track region in a single-frame image is affirmed by Canny edge detection, probabilistic Hough transform, and line segments screening. By adopting the improved MobileNet CNN model, abnormal objects detection and categorization in track region in a single-frame image is carried out. From the result of single-frame detection by the inter-frame correlation optimization in the video, the final results of the video track abnormal objects detection are obtained. Then, a self-built aerial video track region image dataset is used for test. Results demonstrate that the proposed method is suitable for multiple types of abnormal objects in aerial video, with capability of realizing effective detection.