A Multimodal Framework for MODT using Enhanced Correlation-Based Affinity Metrics in Autonomous Driving

3D object tracking plays a pivotal role in $3D$ com-puter vision, with significant applications in robotics, autonomous vehicles, and human-computer interaction. Despite progress, leveraging multimodal information to enhance the accuracy of multi-object detection and tracking (MODT) remains a key research challenge. To address this, we introduce a multimodal multi-object tracking (MOT) framework based on enhanced Affinity computation-based multi-object detection and tracking (ACMODT), specifically designed for autonomous driving scenar-ios. This framework integrates data from cameras and LiDAR sensors to deliver more reliable feature extraction and correlation for real-time tracking and detection. Our approach employs a deep neural network (DNN) that combines image (2D) and point cloud (3D) data for simultaneous object detection, tracking, and association. We design a reliable module to calculate motion and appearance relationships in 3D space while accounting for multiple occlusions and harsh weather conditions. We also create a unified data association module to optimize detection reliability, object associations, and start-end estimation. Experiments taken on the KITTI car tracking dataset and RADIATE dataset demonstrate our method achieves superior tracking accuracy and precision compared to existing approaches.