Towards a Multimodal Loop Closure System for Real-Time Embedded SLAM Applications

Multimodal SLAM algorithms improve its robustness and accuracy in complex and dynamic environments. However, these improvements come at the cost of increased computational requirements. The systemic-level study of the SLAM problem is crucial to designing a practical, stable and versatile solution, adaptable to embedded and real-time systems. We have studied the various processing stages of the system in order to propose contributions to the multimodal loop closure level for SLAM applications, and its computational architecture. This study began with an in-depth analysis of the impact of multimodal information representation on loop closure accuracy and its influence on trajectory drift reduction. We developed a fusion method based on a similarity-guided particle filter, which was evaluated using various dataset. The results obtained showed an improvement in localization’s accuracy. We proposed a heterogeneous architecture model (CPU-GPU and CPU-FPGA) for inter-modal scene descriptor computation. This architecture was able to deliver superior performance in terms of processing time.