Green pepper fruits counting based on improved DeepSort and optimized Yolov5s

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate green pepper yield estimation is vital for harvest and storage planning.
• Detecting green peppers is challenging due to their color similarity to leaves and frequent occlusion.

Purpose of the Study:

• To develop an automatic green pepper fruit counting method using object detection and multi-object tracking.
• To improve the accuracy and efficiency of green pepper yield estimation.

Main Methods:

• A novel CS_YOLOv5s model was designed for green pepper detection, incorporating a Slim-Neck with GSConv and CBAM attention mechanism.
• The DeepSort algorithm was optimized using appearance matching and track optimization from SportsTrack for improved multi-object tracking.
• Performance was evaluated using metrics like mAP, Precision, Recall, Detection Time, MOTA, MOTP, ID switch, ACP, MAE, and RMSE.

Main Results:

• The CS_YOLOv5s model achieved 98.96% mAP, 95% Precision, and 97.3% Recall with a detection time of 6.3 ms, outperforming YOLOv5s.
• Optimized DeepSort reduced ID switches by 29.41% and improved green pepper counting with 95.33% ACP, 3.33 MAE, and 3.74 RMSE.
• The CS_YOLOv5s model demonstrated superior counting accuracy and robustness compared to YOLOv5s when integrated with the optimized DeepSort algorithm.

Conclusions:

• The CS_YOLOv5s model and optimized DeepSort algorithm provide an effective solution for automatic green pepper counting.
• This approach enhances the accuracy and efficiency of agricultural yield estimation, aiding in strategic planning.