MOJO: MOtion Pattern Learning and JOint-Based Fine-Grained Mining for Person Re-Identification Based on 4D LiDAR Point Clouds

Abstract

Person Re-identification (ReID) primarily involves the extraction of discriminative representations derived from morphological characteristics, gait patterns, and related attributes. While camera-based Person ReID methods yield notable results, their reliability diminishes in scenarios involving long distances and limited illumination. LiDAR enables the precise acquisition of human point cloud sequences across extended distances, unaffected by variations in lighting or similar factors. Nevertheless, current LiDAR-based Person ReID techniques are limited to static measurements, rendering them susceptible to perturbations from attire variations, occlusions, and similar confounding factors. To address these issues, this manuscript introduces MOJO, which is applied to 4D LiDAR point clouds to extract unique motion patterns specific to individuals. To characterize the motion patterns across two consecutive point cloud frames, MOJO employs optimal transport to compute point-wise motion vectors, thereby enabling the identification of discriminative implicit motion information. To mitigate the attenuation of point cloud density induced by self-occlusion during dynamic motion, MOJO leverages inverse point-wise flow information to integrate forward frames, thereby yielding a comprehensive representation, whilst concurrently ameliorating the effects of heterogeneous density distribution within localized regions of the 4D point cloud data. Additionally, the inherent unordered nature and sparsity of 4D point clouds present significant obstacles to capturing discriminative features. We develop the 3D joint graph to extract scalable fine-grained traits and employ the joint pyramid pooling module to conduct hierarchical spatiotemporal aggregation across the 4D point clouds. Extensive experimental evaluations demonstrate that MOJO achieves state-of-the-art (SOTA) accuracy on the LReID dataset (for LiDAR-based Person Re-identification) and SUSTech1k dataset (for LiDAR-based Gait Recognition) without any pre-training while exhibiting robust performance across various point cloud densities. Our code will be available at https://github.com/O-VIGIA/MOJO