Multi-object tracking (MOT) is a significant challenge within the field of computer vision, holding various practical applications. Nevertheless, it is still a challenge to extract more discriminative features, more accurate prediction of target motion trajectories and better matching strategies in tracking tasks. This study delves into a large number of multi-target tracking algorithms and makes improvements in feature extraction, motion prediction, similarity calculation and data association. A dual-attention module has been incorporated into the feature extraction network with the aim of enhancing the feature information; a position and velocity information fusion module is designed in the motion prediction stage to learn the long-term dependency of target motion information; and finally, a pedestrian grouping hierarchical data association module is designed to accomplish tracking. Our study shows that we achieve competitive performance on popular multi-object tracking benchmarks (e.g., MOT16, MOT17) compared to current state-of-the-art methods.