A package auto-counting model based on tailored YOLO and DeepSort techniques

In the industrial area, deployment of deep learning models in object detection and tracking are normally too large, also, it requires appropriate trade-offs between speed accuracy. this paper, we present a compressed identification model called Tailored-YOLO (T-YOLO), builds lighter neural network construction based on T-YOLO DeepSort. The greatly reduces number parameters by tailoring two layers Conv BottleneckCSP. We verify realizing package counting during input-output warehouse process. theoretical analysis experimental results show that mean average precision (mAP) is 99.50%, recognition accuracy 95.88%, 99.80%, recall 99.15%. Compared with YOLOv5 combined DeepSort model, proposed optimization method ensures packages 11MB.