| Establishing an intelligent transportation system helps alleviate traffic congestion.Vehicle detection technology is the foundation of intelligent transportation systems.Compared with the traditional vehicle detection methods,the vehicle detection technology based on convolutional neural network(CNN)has the advantages of high accuracy,fast speed,flexible installation and maintenance,wide application range,strong extensibility,convenient management and comprehensive traffic information,so it has a broad application prospect in the field of intelligent transportation.However,running a CNN detection model is often accompanied by a large amount of calculation and memory consumption.To maintain efficient calculation efficiency,there is a large demand for hardware performance,which is greatly limited in engineering applications.Model compression is a useful technique to improve the efficiency of model computation,which makes it possible to deploy solutions in resource-limited scenarios by increasing the speed of model operation by reducing model parameters and computation.This paper investigates cutting-edge algorithms in the field of target detection,and studies the widely used Anchor-based method and the novel keypoint-based method,which are applied to vehicle detection tasks through transfer learning.According to the different characteristics of the above target detection methods,this paper has developed implementation schemes for vehicle detection tasks and optimized them:Scheme 1: Optimization of YOLOv3 algorithm based on Anchor.Aiming at the problem of high model complexity leads to larger parameter redundancy,this paper applies a model compression algorithm to optimize it.Firstly,design a lightweight network replaces the feature extraction network of the model.The design process is divided into three parts: compact model design,knowledge distillation and model pruning.then through Python programming,implement the vehicle detection algorithm model on the Tensor Flow framework and train it through the public vehicle dataset.The improved algorithm achieves a speed increase of more than 2 times and a reduction of nearly 4 times the memory footprint while maintaining the accuracy basically unchanged.Scheme 2: Optimization of the Corner Net-Squeeze algorithm based on key points.In order to solve the problem of corner matching errors when there are many similar(vehicle)targets in a picture,the algorithm is optimized from the model training stage and the post-processing stage,including three parts: training loss function adjustment,non-maximum suppression improvement and post-processing threshold setting.Then train the model using the same environment as scheme 1.The improved algorithm improves the detection accuracy by 0.12 while the speed and memory usage remain unchanged.The scheme 1 optimized model has a faster detection speed of 34 FPS and a detection accuracy of 0.95,which is suitable for real-time detection tasks.The scheme 2 optimized model has a higher detection accuracy of 0.98 and a detection speed of 18 FPS,which is suitable for tasks with high accuracy requirements.Combined with the actual application requirements,the model inference stage was further compressed using INT8 quantization technology,and the vehicle detection model was called through C++ programming.Finally,this paper completed package test functions and produced a rapid vehicle test software development kit(SDK). |
