The construction area of a trailing suction hopper dredger has dense trajectories and separate mud dumping and dredging areas. However，these trajectories have different densities，and traditional trajectory recognition technology fails to effectively recognize the construction behavior and mode of the dredger and thus cannot be successfully applied. In view of these problems，this paper proposes an unsupervised framework for recognizing the construction behavior of the dredger. Firstly，the paper solves the problem of trajectory jump based on the Kalman filter algorithm and improves the quality of trajectory data. Then，the paper uses the HDBSCAN algorithm to identify mud dredging and dumping trajectories with different densities simultaneously and solves the problem of difficult parameter setting by the traditional DBSCAN algorithm in the case of uneven density between classes. Finally，the paper establishes a GMM model based on directional factors，so as to further identify the mud transportation and return trajectories. The results show that the above method can quickly and accurately identify the construction trajectory of a trailing suction hopper dredger.