Underwater target tracking is a fundamental technology for marine development, providing real-time position estimates of the interested targets. However, due to the harsh underwater environment and the noncooperativity of targets, improving tracking accuracy remains a challenge, especially for highly-maneuvering targets. To address this problem, based on multi-autonomous underwater vehicle (multi-AUV) networks, this paper extends the idea of interacting multiple models (IMM) and designs a bearing-only cooperative tracking algorithm in the consideration of the harsh underwater acoustic channels. Specifically, in position prediction, the combination of historical information and the concept of IMM reduces the severe time-lagged effect in traditional prediction methods and the model reliance in standard IMM filters. Then, during position update, a rigidity-assisted relative position representation is designed based solely on bearing measurements, which alleviates the impact of information loss due to communication interruptions, significantly enhancing the continuity of target tracking. Moreover, the algorithm design also considers various uncertainties that may concurrently occur underwater (e.g., error accumulation and model mismatches), and robust optimization strategies with the principle of maximum entropy are designed to enhance the environmental adaptability. Through various simulations and field experiments, the advantages of the proposed method have been validated.