Research On Vehicle Detection Algorithm Based On Lightweight Deep Learning Model

In recent years,with the rapid development of the economy,the number and types of vehicles are increasing,and the management of urban transportation has been under enormous pressure.In order to solve this problem,the concept of intelligent transportation with intelligent transportation systems as the core has emerged,and the core issue is whether to detect vehicle objects and identify vehicle models quickly and accurately.Thanks to the breakthrough of deep learning algorithms,object detection algorithms based on convolutional neural networks can achieve faster and more accurate detection of objects in images,providing technical support for the commercial implementation of vehicle object detection and recognition.The main research content of this article is as follows:Firstly,this paper firstly constructs a dataset called Objects365-vehicle based on the vehicle data from the Objects365,and then expands it into a dataset called Objects21-vehicle using vehicle images collected from the network,and carries out data cleaning and annotation.Specifically,the newly constructed dataset,Objects21-vehicle,has a total of 158431 images and 498955 annotation boxes,and its vehicle types include 6 primary categories and 21 secondary categories.In order to initially evaluate the detection performance of object detection algorithms on different types of vehicles in this dataset and compare the advantages and disadvantages of the two dedicated vehicle datasets constructed in this paper,a comparison experiment was conducted between these two dedicated vehicle datasets based on the YOLOv5s network and its algorithm framework.The experimental results show that the model trained on dataset Objects21-vehicle has significantly improved metrics in most categories compared to the model obtained by training on dataset Objects365-vehicle,which indicates that the dataset Objects21-vehicle has more diversity after data expansion.Secondly,in order to further improve the detection performance of the algorithm on vehicle objects,this paper uses the YOLOv5-s network and its algorithm framework,as well as the model trained on the dataset Objects21-vehicle,as the benchmark model,and conducts comparison experiments on the models trained with different algorithm parameters,network size and algorithm framework,as well as ablation experiments on the models trained based on the improved network structure and its algorithm framework.Specifically,in order to alleviate the problem of missing small object vehicles and enhance the feature extraction ability of occluded object vehicles,this paper adds a small object detection head to the three detection heads at the output of the original network and adds an attention mechanism module to the feature extraction module of the network;at the same time,in order to improve the convergence effect of neural network training,this paper optimizes the boundary box loss function in the original algorithm framework.The experimental results show that compared to the baseline model,the model trained on the dataset Objects21-vehicle and the improved algorithm improves the average accuracy by 2.6%on average at the cost of only a 2.2 ms decrease in the average inference speed per image,effectively improving the detection performance of smaller sized and heavily occluded vehicle objects.Thirdly,in order to realize the lightweight deployment of vehicle object detection model,this paper selects AidLux,a cross-platform application system for deep learning,and Android mobile phones as the edge computing deployment platform,and converts the model trained based on the object detection algorithm from the "pt" format to the "onnx" format and then to the "tflite" format,and completes the deployment and testing.Experiments have shown that the vehicle object detection algorithm can be deployed conveniently on edge device such as mobile phones,and can achieve fast and accurate detection results,which also validates the feasibility of its landing application.