Optimized Deep Convolve Grey Wolf NeuroNet for Robust LiDAR-Based Object Classification in Autonomous Systems

Engineering notes

Experimental results on benchmark Kaggle datasets demonstrate that the proposed model significantly outperforms existing models, achieving a classification accuracy of the proposed DCGWN model, significantly outperforms them by providing a 0.9978, a precision of 0.9969, a recall of 0.9959, an F1 score of 0.9964, and a 0.9910 IoU of 0.9910. OD-CNN, CNN-VFE, DNN-Attn, and Siam-CNN outperformed all the existing models, demonstrating superior robustness under noisy and sparse conditions.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Abstract Accurate object classification with LiDAR data should ensure safe navigation in complex and dynamic environments, as it is a critical function of autonomous vehicles. However, existing deep learning-based LiDAR classifiers may perform poorly because of their susceptibility to the inherent characteristics of LiDAR data, namely, noisy sensors, sparse point clouds, and repeated and redundant feature representations. To address these issues and challenges, in this an optimized deep convolutional grey wolf neural network (DCGWN) that is capable of providing robust LiDAR object classification was developed. First, raw LiDAR point cloud data are processed using several techniques, including noise removal, point separation on the ground, point normalization, and voxelization, to establish an effective input structure. Next, a deep CNN is used to identify the hierarchical spatial features of the LiDAR point clouds, which are converted into compact feature vectors. The grey wolf optimizer (GWO) is applied to select optimal features by truck eliminating redundancy and enhancing discriminative capability. Finally, a deep neural network (DNN) performs accurate classification using optimized features. Experimental results on benchmark Kaggle datasets demonstrate that the proposed model significantly outperforms existing models, achieving a classification accuracy of the proposed DCGWN model, significantly outperforms them by providing a 0.9978, a precision of 0.9969, a recall of 0.9959, an F1 score of 0.9964, and a 0.9910 IoU of 0.9910. OD-CNN, CNN-VFE, DNN-Attn, and Siam-CNN outperformed all the existing models, demonstrating superior robustness under noisy and sparse conditions. Therefore, the proposed DCGWN framework provides an efficient and reliable method for real-time autonomous driving applications.

Links and sources

Comments