The size and mass of fish are crucial parameters in aquaculture management. However, existing research primarily focuses on conducting fish size and mass estimation under ideal conditions, which limits its application in actual aquaculture scenarios with complex water quality and fluctuating lighting. A non-contact size and mass measurement framework is proposed for complex underwater environments, which integrates the improved FishKP-YOLOv11 module based on YOLOv11, stereo vision technology, and a Random Forest model. This framework fuses the detected 2D key points with binocular stereo technology to reconstruct the 3D key point coordinates. Fish size is computed based on these 3D key points, and a Random Forest model establishes a mapping relationship between size and mass. For validating the performance of the framework, a self-constructed grass carp dataset for key point detection is established. The experimental results indicate that the mean average precision (mAP) of FishKP-YOLOv11 surpasses that of diverse versions of YOLOv5–YOLOv12. The mean absolute errors (MAEs) for length and width estimations are 0.35 cm and 0.10 cm, respectively. The MAE for mass estimations is 2.7 g. Therefore, the proposed framework is well suited for application in actual breeding environments.

Accurate monitoring of fish movement is essential for understanding behavioral patterns and group dynamics in aquaculture systems. Underwater scenes—characterized by dense populations, frequent occlusions, non-rigid body motion, and visually similar appearances—present substantial challenges for conventional multi-object tracking methods. We propose an improved CenterTrack-based framework tailored for multi-fish tracking in such environments. The framework integrates three complementary components: a multi-branch feature extractor that enhances discrimination among visually similar individuals, occlusion-aware output heads that estimate visibility states, and a three-stage cascade association module that improves trajectory continuity under abrupt motion and occlusions. To support systematic evaluation, we introduce a self-built dataset named Multi-Fish 25 (MF25), continuous video sequences of 75 individually annotated fish recorded in aquaculture tanks. The experimental results on MF25 show that the proposed method achieves an IDF1 of 82.5%, MOTA of 85.8%, and IDP of 84.7%. Although this study focuses on tracking performance rather than biological analysis, the produced high-quality trajectories form a solid basis for subsequent behavioral studies. The framework’s modular design and computational efficiency make it suitable for practical, online tracking in aquaculture scenarios.