Abstract: Destination prediction of moving objects is an important part of location-based service.There are always some difficult problems in this field,such as sparse data and long-term dependence.In order to solve these problems effectively,firstly,a trajectory segmentation method based on the minimum description length strategy (MDL) is introduced,which can obtain the best tra-jectory segmentation,improve the similarity between tracks and realize the simplification of trajectories.Then,the segmented data are processed by image processing and local extraction,and the trajectory destination is clustered to add labels to the trajectory data.Finally,this paper proposes a deep learning framework CNN-LSTM based on convolution and long-short term memory.In this framework,local image data and labels firstly are taken as the input of the CNN model,and the effective information is preserved through the depth extraction of spatial features.Then,the LSTM algorithm is used for training and destination prediction.Extensive experiments are carried out on real trajectory dataset of moving objects .The results demonstrate that the CNN-LSTM method proposed in this paper has a strong learning ability and can better capture the spatiotemporal correlation of trajectories.In comparison to state-of-the-art and latest prediction algorithms,this method has high accuracy of destination prediction.

Key words: CNN, Destination prediction, LSTM, Moving objects, Trajectory