Research And Application Of Mobile Robot Real-time Object Detection Algorithm Based On Deep Learning

Object deteceion is the classification and location research of target object,and it is one of the most important and challenging branches of computer vision,which has become a hot-spot research in many fields.In recent years,the object detection technology based on deep learning has made great progress because of its remarkable advantages,which has appeared many excellent algorithms and articles.Object detection is widely used in visual image analysis,intelligent monitoring,UAV,industrial detection,biomedical instruments,automatic driving and other fields,with very important practical significance.At present,the object detection algorithm based on deep learning often needs to run on the equipment with strong computing power because of its complex network structure and large amount of computation.Therefore,in the case of limited computing resources,the object detection algorithm is difficult to integrate into the embedded system.Mobile robot object detection is a technology of classification and location for object,which is used for robot real-time perception target object of surrounding environment in the movement,which can help the robot to effectively avoid obstacles and ensure the safety of driving.Most of the mobile robots object detection use traditional detection algorithms,which need to extract features manually.There are some disadvantages,such as poor generalization ability,poor robustness,low accuracy and low recall rate,which make it difficult to use in complex scenes.The precision of object detection based on deep learning is far more than that of traditional object detection methods,but the mobile robot object detection technology based on deep learning basically transmits the image obtained by the camera to a high-performance computer system through the network for image processing.Due to the limited computing power of the mobile robot embedded platform,the object detection model based on deep learning is not suitable for running on the embedded device,which makes it impossible to build the autonomous mobile robot with complex artificial intelligence system.In order to solve these problems,this paper proposes an improved algorithm to realize real-time object detection on the embedded platform of mobile robot.The algorithm can sense the surrounding environment in real time to help the mobile robot navigate autonomously and avoid obstacles automatically.We analyze the most advanced object detection algorithm,and choose the Mobile Net-SSD network framework as the solution.In this paper,a series of improvement measures are taken for Mobile Net-SSD,including improving the design of default box,replacing feature extraction network with Mobile Net V2 and model quantization,to adapt to the specific platform,and achieving the balance of speed and precision on the embedded platform based on Jeston TX2.The model is evaluated by mean Average Precision(m AP)and frames per second(FPS),experimental results show that our object detection algorithm achieves the speed of 22.43 FPS(including all steps)and the detection precision of 69.83% m AP on the mobile robot embedded platform,realizing the function of detecting target objects in the surrounding environment of the mobile robot in real time,which is of great significance for autonomous navigation and automatic obstacle avoidance of mobile robot.At the same time,the research lays a foundation for the application of object detection technology based on deep learning in the mobile robot embedded platform,and solves similar technical scheme problems.