To be successfully used in daily life situations, exoskeletons should be effective across multimodal scenarios, including walking on various terrains, and transitions between locomotion modes such as walking-to-stop. Correct continuous gait phase estimation is essential for high-level walking assistance control. Despite the impressive advances in gait phase estimation for a variety of locomotion modes, transition gait phase estimation is rarely researched, leading to the jittering of exoskeletons during walking-to-stop transitions. We propose an optimized phase oscillator (PO-opt) that estimates the gait phase correctly during transition gaits in multimodal locomotion, which is beneficial to eliminating the jittering. In the phase plane, a lateral axis extreme difference (LAED) is adopted to classify transition gaits. The threshold of LAED for transition gaits in multimodal locomotion was preliminarily determined by simulation, which was then applied and validated in experiments. Simulation results indicated that a threshold of 15.0 was suitable for transition gaits classification during treadmill walking, free walking, and ramp ascent/descent, while results of the experiment showed that a threshold between 6.5 and 10.5 was applicable for treadmill walking, free walking, and stair ascent/descent. In particular, the jittering elimination rates for 3, 4, and 5 km/h treadmill walking were improved from 29%, 21%, and 4% (PO model) to 100%, respectively, when the threshold of LAED was set at 15.0 in PO-opt model. The results indicated a significant increase in the rate of jittering elimination when the PO-opt model was applied. The model holds great promise in real-world applications for prostheses and other types of exoskeletons.