A Multi-scale Deep LSTM Model for 3D Point Cloud Classification

3D point cloud classification has become a fundamental research topic across multimedia and intelligent communication fields, playing a critical role in applications such as autonomous driving, smart cities, and immersive media experiences. Due to the inherent complexity of 3D data, traditional approaches often transform point clouds into standard formats or design complex neural network architectures for processing, which may result in information loss and computational inefficiency. This paper proposes an innovative multi-scale serialization method for 3D point cloud data, aiming to address the challenges posed by the inherent complexity of point clouds and the information loss caused by local feature aggregation. The primary contribution of this work is the introduction of the Sequencer module, which effectively captures dependencies among serialized point cloud data to address the challenges associated with point cloud properties in feature learning. Additionally, we adopt a random sampling strategy to generate point cloud sequences with varying numbers of points as model input, expanding the model’s receptive field and ensuring a more comprehensive feature learning process. This approach aligns with the demands of next-generation broadcasting and networking systems for robust and adaptive multimedia signal processing. Experimental results on the ModelNet40 and ScanObjectNN datasets demonstrate that the proposed method achieves competitive performance and exhibits strong generalization capabilities across diverse scenarios, further validating its effectiveness.