Research And Application Of Object Detection Based On Embedded Platform In Roadside Parking Scene

With the continuous development of the field of computer vision,the object detection technology is widely used in the intelligent parking industry,but there are many vehicle detection problems when the universal object detection algorithm is applied to the specific roadside parking scene.In order to solve the above problems,this thesis improves the YOLOv5 s algorithm to obtain a vehicle detection algorithm suitable for this scene,which improves the vehicle detection performance of this scene.Firstly,the vehicle data set is analyzed and processed,and the CTB vehicles of the COCO dataset and the private roadside parking data set are combined to form the CTB vehicle training set,the roadside parking test set and the COCO vehicle test set.Considering that there are fewer small target vehicles in this scene,a new COCO vehicle test set is formed after deleting the vehicle data with less than 32×32 pixels in the COCO test set,and a CTBT vehicle data set containing tricycles is also produced.Then,a YOLOv5s-CTB vehicle detection algorithm suitable for this scene is constructed.Considering the inaccurate positioning of the vehicle boundary box in this scenario,the SIOU loss function is used to replace the original border regression loss function.Considering the imbalance of vehicle feature fusion,the BiFPN network is used to optimize the original feature fusion network.In order to improve the inference speed of the algorithm,the GSConv lightweight convolution module is used to optimize the network structure of YOLOv5 s.Finally,the YOLOv5s-CTB vehicle detection algorithm constructed by combining the above improvements improves the mAP by 2.3% on the roadside parking test set,by 0.7% on the COCO vehicle test set,and by 0.8% on the new COCO vehicle test set.In addition,this thesis also adds the detection of tricycles in this scene,constructs a YOLOv5s-CTBT vehicle detection algorithm suitable for this scene,and analyzes the CTBT vehicle detection problem generated by the introduction of tricycles in this scene.In response to the above problems,the prediction layer of YOLOv5 s is improved,the coupling detection head is replaced by the decoupling detection head,and the decoupling detection head is designed in a lightweight way.Due to the small number of sample instances in fine-tuning training,in order to improve the feature extraction ability of the network,the RepVGG module and C2 f module are used to improve the backbone network,and the shallow network structure in the backbone network is frozen.Finally,the constructed YOLOv5s-CTBT vehicle detection algorithm improves the mAP by 2%,the tricycle by 2.6%AP,and the inference speed by 8.5% on the CTBT vehicle test set.Finally,the above improved vehicle detection algorithm was deployed to the Jetson TX2 embedded platform,and Tensor RT technology was used to accelerate the reasoning of the improved vehicle detection algorithm.The improved YOLOv5s-CTBT vehicle detection algorithm is also used to improve the roadside parking management system,and the system accuracy is improved by 12.2%.