CB-YOLO-DeepSORT : Real-Time Yield Estimation for Tomatoes in Greenhouses

To address the challenge of maturity monitoring and fruit quantity estimation in tomato yield prediction under greenhouse conditions, this paper proposes the CB- YOLO-DeepSORT framework for continuous tracking and dynamic yield estimation. Built upon YOLOv5s, the framework incorporates the CBAM attention mechanism to enhance tomato maturity recognition in occluded environments and integrates the CIo U loss function to improve localization accuracy, ensuring precise ripeness mapping for reliable yield estimation. Then, the DeepSORT -based multi-object tracking algorithm is applied to continuously monitor tomatoes at different maturity stages, maintaining identity consistency to enhance the stability of real-time tomato yield estimation. Validation confirms the enhanced CB- YOLO attains 99.5% mAP@O.5 on test data, surpassing YOLOv5s by 2.5 percentage points. The proposed system achieved MOTA and MOTP scores of 82.350/0 and 84.5% in medium-density scenarios, demonstrating robust tracking performance under moderate target congestion. Finally, a web-based monitoring system deploys autonomous robots with depth cameras to track tomato maturity in greenhouses. Real-time data is visualized through daily snapshots (pie charts and counts of ripe/unripe fruits) and weekly maturation trends (line charts and histograms), providing actionable insights to optimize harvest scheduling and resource allocation for precision agriculture.