Towards an embedded architecture based back-end processing for AGV SLAM applications

Place recognition plays a crucial role in the Simultaneous Localization and Mapping (SLAM) process of self-driving cars. Over time, motion estimation is prone to accumulating errors, leading to drift. The ability to accurately recognize previously visited areas through the place recognition system allows for the correction of these drift errors in real-time. Recognizing places based on the structural aspects of the environment tends to be more resilient against variations in lighting, which can cause incorrect identifications when using feature-based descriptors. Nevertheless, research has predominantly focused on using depth sensors for this purpose. Inspired by a LiDAR-based approach, we introduce an inter-modal geometric descriptor that leverages the structural information obtained through a stereo camera. Using this descriptor, we can achieve real-time place recognition by focusing on the structural appearance of the scene derived from a 3D vision system. Our experiments on the KITTI dataset and our self-collected dataset show that the proposed approach is comparable to state-of-the-art methods, all while being low-cost. We studied the algorithm’s complexity to propose an optimized parallelization on GPU and FPGA architectures. Performance evaluation on different hardware (Jetson AGX Xavier and Arria 10 SoC) shows that the real-time requirements of an embedded system are met. Compared to a CPU implementation, processing times showed a speed-up between 4x and 10x, depending on the architecture.