Fruit occlusion, dense clustering, platform vibration, and cross-frame re-identification errors reduce the reliability of video-stream yield estimation in high-density dwarf apple orchards. To address these limitations, this study developed a vision-based framework for apple detection, tracking, dynamic counting, and yield estimation. Apple detection was improved by embedding a Shuffle Attention (SA) module into YOLOv12n and introducing SlideLoss for hard-sample reweighting, yielding the YOLOv12n-SA-SlideLoss detector. Following detection, the DeepSORT tracking module was also improved by augmenting it with global motion compensation (GMC) and a dual-criteria association strategy based on motion consistency and appearance similarity. Yield estimation was formulated as a counting-based compensation model using dynamic fruit counts, mean fruit mass, and an occlusion compensation factor, and was compared with a pixel-geometric regression model based on detection boxes. Field experiments were conducted in a high-density dwarf ‘Fuji’ apple orchard in Aksu, Xinjiang, China. YOLOv12n-SA-SlideLoss achieved an mAP@0.5 of 95.7%, exceeding the YOLOv12n baseline by 2.3%. The improved DeepSORT reached 87.4% multiple object tracking accuracy and 86.7% multiple object tracking precision, while reducing identification switches by 22.2% relative to the original algorithm. The dynamic counting framework increased average counting accuracy from 85.6% to 93.5%. For yield estimation, the counting-based compensation model achieved a mean prediction accuracy of 85.2%, an RMSE of 18.5 kg, and a mean bias of −17.2 kg across six validation samples (n = 6), each covering multiple trees. Overall, the results indicate that combining detection, motion-compensated tracking, and counting-based correction improves stability of single-view video-stream yield estimation in high-density dwarf apple orchards.