Deadlocks are a common defect of parallel programs. To detect deadlocks, dynamic deadlock analysis methods build models such as lock graphs and segment graphs according to program running traces. However, a lock graph and its existing variants cannot distinguish different executions of one lock acquisition statement in a loop structure. The lock set used in extended lock graphs cannot capture those locks which were once held and then released. A segmentation graph cannot model the inter-segment dependencies caused by the coupling of lock release/acquisition operation and thread start operation. The above problems have led to a variety of false positives. To address this issue, existing lock graph and segment graph models are improved. Specifically, a lock graph is extended with statement execution order information. A segmentation graph is expanded with lock acquisition and release information. Furthermore, segments in a segmentation graph are more finely divided in the improved model to capture the inter-segment dependencies caused by lock objects. Eventually, based on the improved models, a new deadlock detection method is proposed. It can eliminate the aforementioned false alarms effectively and improve deadlock detection accuracy. A corresponding prototype system is developed for experimental evaluation. The validity of the proposed method is verified through experiments.