Yang Gao

Recognizing 3D objects from point clouds is a crucial technology for autonomous vehicles. Nevertheless, LiDAR (Light Detection and Ranging) point clouds are generally sparse, and they provide limited contextual information, resulting in unsatisfactory recognition performance for distant or small objects. Consequently, this article proposes an object recognition algorithm named Adaptive Scale and Correlative Attention PointPillars (ASCA-PointPillars) to address this problem. Firstly, an innovative adaptive scale pillars (ASP) encoding method is proposed, which encodes point clouds using pillars of varying sizes. Secondly, ASCA-PointPillars introduces a feature enhancement mechanism called correlative point attention (CPA) to enhance the feature associations within each pillar. Additionally, a data augmentation algorithm called random sampling data augmentation (RS-Aug) is proposed to solve the class imbalance problem. The experimental results on the KITTI 3D object dataset demonstrate that the proposed ASCA-PointPillars algorithm significantly boosts the recognition performance and RS-Aug effectively enhances the training effects on an imbalanced dataset.

As a basis of many missions, the accuracy of localization is highly important for mobile robots. For the generally used map matching based localization algorithms, the accuracy of localization, which is described by localizability, is greatly impacted by the environment. Consequently, this paper proposed a novel method to predict the localizability for the map matching based localization algorithms, based on the environment map. Firstly, the uncertainty of localization in map matching and dead-reckoning is analyzed based on which entropy of localization is chosen to describe the localizability instead of the generally used covariance. Next, based upon the flow chart of the map-based localization algorithm, a localizability predictor, which is composed of three different models, is designed to predict the entropy. Here a Convolutional Neural Network (CNN) is designed for the first model to predict the entropy of localization that comes from map matching. A Long Short-Term Memory (LSTM) neural network is designed for the second model to predict the entropy that comes from the dead-reckoning. Finally, a Multilayer fully connected Neural Network (MNN) is designed for the last model to predict the entropy after fusing the entropy results that come from the two models described above. Both simulation results and experimental results have proven that the proposed predictor can offer a better estimator of localizability compared to other existing approaches.

As a fundamental capability of mobile robots, path planning highly relies on the accurate localization of the robot. However, limited consideration for the localizability (which describes the capability of acquiring accurate localization) has been made in path planning. This brings a high risk of choosing a path that is optimal but results in the robot easily getting lost. There exist two key challenges to address this problem: 1) How to evaluate the localizability of a path and its impact on path planning. 2) How to balance the localizability of a path and the standard path planning criteria (e.g., shortest travel distance, obstacle-free path, etc.. To overcome the two challenges a new path evaluation method is required. So we first analyzed the uncertainty that comes from dead-reckoning and map matching. Then the localizability was estimated by the fusion of the uncertainty coming from both of them. Based on that, the impact of the localizability on the path planning task has been evaluated by an evaluation function. By combining the localizability evaluation function with traditional criteria (e.g., shortest length, obstacle-free path, etc.), a new path evaluation function for path planning is established. Both simulation and experimental studies show that the new path evaluation function can offer a balance between the localizability and the traditional criteria for path planning.