|
Aiming at the problem that plums detection in a natural environment is subject to serious environmental interference and the detection method is not easy to be deployed on mobile devices, a target detection method suitable for jetson nano terminal is proposed, which can accurately detect plums and make the model adapt to the needs of the mobile terminal. A total of 1000 ripe plum images were collected and 20 images from each typical picking scene were selected as the test set. The remaining images are divided into training and validation sets according to 8:2. The YOLOv3 model is modified to accommodate mobile terminals, the main neural network of YOLOv3 is replaced by mobile_v2, and the structure of the FPN is simplified to achieve network compression and improve detection speed. The improved model was trained using the PyTorch framework, and the trained model was converted to an ONNX file, which was moved to the jetson nano. On the jetson nano side, the TensorRT framework is used to parse the ONNX files, generate the model inference engine, and implement model acceleration. The experimental results show that the detection accuracy of the improved YOLOv3 on the test set is 91.27%, and the accuracy of the improved YOLOv3 is 97.85%, 98.20%, 94.78%, 81.66%, and 85.33% under the conditions of slight interference, branches and leaves occlusion, fruit overlap, occlusion and overlap, and insufficient light, respectively. In the experiment, the detection speed is 146FPS for the self-built server and 6FPS for the jetson nano. Experimental results show that the proposed method can meet the accuracy requirements of plum detection in picking scenarios, and the deployment and acceleration of the model on small devices can be achieved, thus laying the foundation for the practical application of automatic picking.
|
