Automatic Unloading System Based On Binocular Vision

With the continuous popularization and development of the Internet,the scale of online shopping transactions continues to expand,logistics has become an integral part of e-commerce.People also put forward higher and higher requirements for logistics.Handling is an important part of logistics.Most enterprises are still in the stage of manual handling and sorting,leading to that the handling become the more complex and low efficiency work in the logistics,which will undoubtedly consume a large amount of manpower.In recent years,robot technology has gradually matured.The application of intelligent robot in logistics industry,aiming at the loading and unloading system,can reduce the manpower,realize precise control and achieve continuous work.Intelligent transportation and handling system combined with vision sensor and robot arm has become the main research direction.The purpose of this paper is to design an automatic unloading system based on binocular vision to realize the automation of handling and unloading.According to the overall design requirements of the system,the function of the system is divided into two parts:hardware and software.This design uses Xilinx ZYBO^TM development board,robot arm,binocular camera and touch sensors to build the basic hardware structure of the system.This paper proposes two kinds of software implementation:the one is multi-target recognition algorithm based on local features,another is multi-target recognition algorithm based on full convolution network.In the multi-target recognition algorithm based on local features,the SIFT algorithm is used to extract the local features of the image.Then,the local features are combined with the depth information and the location information,and the Mean Shift is used to cluster.The target recognition is achieved at the same time of segmentation.The experimental results show that the algorithm has better segmentation and recognition effect in the scene with complex color information.In the multi-target recognition algorithm based on full convolution network,the VGG16 pre training model is fine-tuned by using data set,and better segmentation and recognition results are obtained.To reduce the long running time of the algorithm,the hardware acceleration of the algorithm is implemented by using FPGA.Experimental results show that,compared with software implementation,the hardware implementation can greatly improve the speed of full convolution network algorithm.