Design And Implementation Of Embedded Object Detection Platform

With the improvement of deep learning object detection,relevant technologies are gradually empowering embedded devices.It provides stronger computing capabilities.Deep learning environment deploying and object detection model training requires expensive computing resources.It is difficult for embedded programmers to study in deep learning by themselves.At present,embedded object detection is gradually applied in public transport operation monitoring management.Therefore,in order to reduce the difficulty to apply deep learning target detection technology,this paper presents a platform design for embedded object detection scene and then provide an embedded object detection platform with strong versatility and simple operation which can apply to the scenes of bus monitoring and management.The platform has the ability of both equipment management and obejct detection model training.The main research in this paper include:1.Designs an embedded target detection platform suitable for transport operation monitoring managementThe platform has both equipment(vehicle-mounted,road testing and other equipment,the same below)management ability and deep learning training and testing ability,which can be applied to embedded target detection scenarios.The platform has the ability to push models,configuration files,parameters online,register the device quickly,to display target detection results in video and text form;to train and test model in a simplified,unified process.2.Achieves each functional module of the platform and verify the platform in the actual sceneFirstly,the platform is mainly divided into four modules:central server,embedded end,file server,forwarding server.The central server mainly provides the relevant services and the ability to interact with users.The embedded end mainly provides the target detection ability and the ability of receiving commands and uploading detection results.The file server mainly provides the model and configuration file storage and management capabilities;The forwarding server mainly provides the ability of forwarding streaming video.Then,this paper does the functional tests and module tests to prove the usability and stability of the relevant services.Finally,this paper tests the platform in the scene of bus station monitoring scene and congestion detection scene to prove that the platform can apply to transport operation monitoring management scene.3.Designs a cascading feature detection algorithm to solve the problem of congestion detection in busFirstly,this paper designs a cascade feature detection algorithm to evaluate the congestion degree in bus.The algorithm obtains the key targets by the target detection algorithm,converts them into mask images by the intermediate connecting blocks and obtains the congestion evaluation results by the classification network.Then,this paper does the experiment to select the optimal detection algorithm.Yolov3 is better than other algorithms with 67.9%accuracy and 68.0 fps,considering the accuracy and speed.This paper does the experiment to select the optimal classification network.The accuracy of AlexNet is better than other networks with 82.7%accuracy and 512.90 fps,considering the accuracy and speed.Finally,this paper select Yolov3 and AlexNet as the detection algorithm and classification network in the cascade feature detection algorithm.