The Recognition Scheme And Implementation Of Kitchen Overalls Based On Hi3559A

Artificial intelligence has good performance in the field of target detection.However,there are two major difficulties in getting the algorithms off the ground.Firstly,the performance improvement of deep learning networks is usually dependent on the huge computing power of the server,which makes it difficult for the model to run on edge side and embedded devices,greatly limiting the application scenarios and the usage cost of the algorithm.Secondly,cloud solutions needs to transmit a large amount of data.This scheme has high time delay and takes up local bandwidth,which increase the cost of the system.To address above problems,we propose a workwear identification solution which can be deployed and inferred on embedded devices,The main works of this research are as follows.1.The project carried on the theoretical analysis of the relevant knowledge and the feasibility of the solution,identified the data handling process,and the basic structure of the system.We also compared the advantages and disadvantages of various feasible solutions.2.By introducing a pipeline processing architecture and using a series of acceleration methods such as model slicing,data quantization,ring caching and parallel filtering,the network inference speed is theoretically increased by a factor of 2.5.3.The project designed a set of embedded master control system with Hi3559 A as the main processor,the uniform recognition algorithm was successfully run on the embedded device,through the design and optimization of the system logic,multi-core deployment scheme,video encoding and decoding and communication protocols.4.This project esigned a set of computer-side inspection consoles,which can realize the device management and the display of network identification results.It also can monitor the system working status in real time.The olution of this project has been tested on Hi3559 A processor,which can achieve the recognition speed of 28 frames and power consumption of only 5.5 W.The complexity of the network model is ensured while the real-time performance of the algorithm is greatly improved.