In autonomous driving, moving object segmentation plays an essential role in improving the accuracy of mapping and path planning. However, accurately and consistently differentiating between objects that are actually moving and those that are potentially moving is still challenging. This paper proposes DeepMHT, a novel approach that combines a multiple hypothesis tracking algorithm (MHT) with deep learning to solve the issue of moving object segmentation of LiDAR point clouds. This paper designs a data association network to extract deep features from the object point cloud and 3D box and fuses these two features to obtain affinity weights. We propose an innovative local hypothesis update model that integrates affinity weights with the weights of MHT to make data associations more accurate and stable. Unlike other moving object segmentation algorithms that calculate differences in the point clouds of adjacent frames, DeepMHT directly estimates the motion state of objects and segments moving objects by their position and velocity, which significantly improves the precision of segmentation. This paper compares DeepMHT with several other moving object segmentation algorithms on the KITTI tracking dataset and finds that DeepMHT is highly accurate.