GAN based simultaneous localization and mapping framework in dynamic environment

Abstract

ObjectiveBecause of interference from dynamic objects, the traditional simultaneous localization and mapping (SLAM) framework execute poorly while operational in a dynamic environment. The Dynamic-SLAM model is introduced by considering the merits of deep learning in object discovery. At the semantic level, the dynamic objects in the new detection are detected to construct the prior knowledge with an SSD object detector. Among the radar and the LiDAR, the synchronization and conversion calibration makes the multi-sensor fusion. Localizing using a camera in a dynamic environment is more challenging because the localization process takes place in moving segments. This further results in unstable pose estimation. A large amount of information of the visual point cloud and high precision of the laser radar information enhances the accuracy of real-time positioning thereby attaining grid map and 3-D point cloud map. MethodsHere, we have proposed Generative Adversarial Network (GAN) with Aquila Optimizer for moving object detection. The LiDAR measurements check the Radar outcomes. The targeted moving objects are determined via Doppler velocity from the radar and their exact location and mass can be estimated with LiDAR and the proposed GAN-AO approach. Hence the GAN-based AO approach is used to segment the objects inside the point clouds. The arrangements of point clouds are made in a particular range to multiply the vertical points with the laser channel number. If the same objects are identified then the angles are analyzed among the image vectors and then labeled identically for the same points of the object. In addition to this, Velocity compensation is made to estimate the actual moving target from the world frame. This is due to the fact that the estimated velocity of the mmW-radar is in the radial direction with respect to the touching sensors. ResultsThe investigation is conducted between different state-of-art methods like Deep Learning (DL), Generative Adversarial Network (GAN), artificial neural network (ANN), Deep neural network (DNN) and propose methods. From this analysis, the proposed method provided 93.8% detection accuracy than other existing methods like DL, ANN, GAN and DNN respectively. ConclusionsWhile comparing to the state-of-art techniques, the proposed method demonstrated superior performance results in terms of tracking, detection, root mean square error (RMSE) and accuracy.