Research On Target Matching And Positioning Based On Deep Learning

In recent years,with the improvement of computing power and the rapid development of deep learning,object detection is one of the hottest research areas in computer vision.Intelligent target detection is not only in the civil fields such as intelligent transportation and automatic driving,also it is of great significance in military field.Target detection and recognition of intelligence is one of the core technologies that determine the combat effectiveness of missile weapon system,and it is also the bottleneck that restricts the development of missile weapon system informatization for a long time.Now the target detection and recognition method based on template matching is mainly adopted for missile models,its theoretical framework is simple,fast calculation speed,can satisfy the general requirement for industrial application.However,the template matching method has poor adaptability to the change of the target’s view field,complex background and interference,and it is difficult to meet the actual combat requirements of intelligent missile weapons.In view of the above problem,this paper designed a target matching and positioning system based on deep learning.The system adopts the target matching network algorithm based on Siamese.Although the algorithm is mainly composed of CNN(Convolutional Neural Networks),Siamese dose not directly output categories,but calculates the similarity of two images through the output vector.It does not require the large sample training of other target detection algorithms,so it is very suitable for transplantation to small mobile devices with limited computing and storage resources.The architecture design of the algorithm hardware acceleration system is similar to TTA(Transport Triggered Architecture),which can control the operation flow of the system by compiling specific instruction program to configure the parameters of the system,so that the processing system can flexibly realize the accelerated operation of the convolutional neural network with different structures.In the core of the system in parallel arithmetic circuit design,256 operation unit array structure with parallel computing,and by using the design method of Time Division Multiplexing structure to adapt to a larger network.In the system storage space design,this paper adopts the storage space dynamic address design technology,the required storage space only occupies 1/64 of the traditional fixed address storage space.Such a design method not only improves the computing speed and efficiency,but also makes the system have a strong commonality.At the same time,it also saves a lot of hardware resources,making it possible to transplant the target matching and positioning system algorithm based on deep learning to the FPGA with limited single chip resources.The specific circuit of the hardware system is designed with Verilog HDL,and 8bit fixed-point data is used for calculation and storage,which can fully meet the accuracy requirements of data calculation.The entire system has been tested on XILINX’s virtex-7 XC7VX485 T with an accuracy of 87.2%,less than 1% lower than the 88.1% accuracy on the GPU.At the same time,it only takes 4.624 ms for the system to process a 128×128 image,the data occupies only 6.6mb of storage space,and the total power consumption is 0.528 w.The system fully meets the requirements of processing time index 10 ms and data storage space index 10 Mb.In general,this paper designed the target matching and positioning system based on deep learning has the advantages of fast computation speed,high recognition rate,less resource occupation,low power consumption,etc.The design ideas of the system’s overall architecture,circuit parallelization design and memory space address design can provide some reference for future deep learning algorithm transplantation in resource-constrained environments.