<p>Underwater target bearing estimation and trajectory tracking are fundamental capabilities of active sonar systems. However, conventional beamforming and model-driven tracking methods often exhibit degraded accuracy and temporal instability in complex underwater environments affected by reverberation, multipath propagation, and non-stationary interference. This paper proposes a two-stage hybrid framework that couples closed-loop adaptive beamforming with detection-driven tracking on bearing images. In the first stage, a radial basis function (RBF) neural network is integrated with an incremental proportional–integral–derivative (PID) controller to form an online closed-loop correction mechanism, producing stable, high-resolution bearing images and improving frame-to-frame bearing consistency. In the second stage, the bearing image is treated as a single-channel acoustic image and processed using a CenterNet detector together with the observation-centric SORT (OC-SORT) association algorithm to track target trajectories across frames. Simulation results demonstrate improved bearing accuracy and enhanced temporal stability compared with representative baseline methods. Field experiments conducted at Qiandao Lake further validate the feasibility of the proposed approach, yielding clear bearing imagery and reliable trajectory tracking under practical measurement conditions. From a system perspective, the proposed pipeline is designed for frame-wise streaming processing, where covariance estimation, dense angular scanning, and deep detection/tracking inference constitute computational hot spots that naturally benefit from parallel acceleration on multi-core GPUs.</p>