A method for realizing high-precision vehicle recognition based on Loongson 2k1000 processor

Abstract

The invention relates to a method for realizing high-precision vehicle recognition based on a Loongson 2K1000 processor, belonging to the technical field of embedded image processing, including vehicle data collection and labeling processing, making deep learning data sets; training deep learning models; The learning model is optimized; the deep learning model is deployed; the video stream is selected for vehicle detection. The present invention realizes the AI ​​application deployment of vehicle detection and recognition on the Loongson 2K1000 processor by cutting and quantizing the model, calls the tmfile file on the 2K1000 processor; verifies for the first time that the 2K1000 processor supports yolo operators; adopts overall discarding for the feature module The method optimizes the performance of the model and improves the efficiency of vehicle detection and recognition; the invention has low cost, is suitable for technology update and replacement, is easy to deploy, and can be used for image recognition applications in other fields by making different data sets.

Description

technical field [0001] The invention relates to a method for realizing high-precision vehicle recognition based on a Loongson 2K1000 processor, which belongs to the technical field of embedded image processing. Background technique [0002] With the increasing sales of automobiles in China, traffic jams, speeding collisions and other accidents occur frequently, the demand for real-time traffic information is naturally derived, in order to allow traffic commanders or participants to obtain necessary traffic information in a timely manner , it is necessary to automatically describe the relevant traffic information in the form of labels, so that the computer can automatically obtain the information and process it into information that is easy for people to understand. [0003] At present, the research on image recognition technology in computer applications has achieved fruitful results. However, there are few research results on image recognition based on domestic chips and em...

AI Technical Summary

Problems solved by technology

[0004] However, the current vehicle recognition accuracy of similar products is not high. Using the convolutional neural network (CNN) as a deep learning model requires a large amount of calculations. There are problems with the model being too large and the detection time being too long. The computing power of the Godson 2K1000 processor is not enough to support short data Time is effectively processed, and deep learning has the defect of being too dependent on GPU

Images