A Lightweight Semantic Understanding Framework for Real-Time Navigation of Mobile Robots in Complex Dynamic Environments

Real-time autonomous navigation for mobile robots operating within complex dynamic environments necessitates a precise, high-speed semantic understanding of their surroundings. This capability is crucial to accurately distinguish between traversable paths and unpredictable dynamic agents. However, existing semantic segmentation frameworks frequently encounter a severe trade-off between computational latency and categorical accuracy. Furthermore, these conventional models often fail to maintain temporal consistency across consecutive video frames, which inevitably leads to erratic navigational behavior and compromised safety. To address these critical limitations, this paper proposes a novel, lightweight semantic understanding framework featuring a Bilateral Asymmetric Encoder-Decoder (BAED) architecture coupled with a specialized Temporal Consistency Module (TCM). The proposed BAED effectively reduces feature redundancy and computational overhead through the implementation of asymmetric spatial-context pathways. Concurrently, the TCM utilizes advanced motion-aware alignment techniques to significantly stabilize semantic predictions over time. Comprehensive experimental evaluations conducted on an NVIDIA Jetson AGX Orin edge computing platform demonstrate that the proposed framework achieves a highly competitive Mean Intersection over Union (mIoU) of 74.2% ± 0.4% while operating at an impressive 45.1 ± 1.2 frames per second (FPS). Compared to the established BiSeNetV2 baseline, the proposed method drastically reduces the total parameter count to merely 1.6M and decreases temporal variance to 0.024 ± 0.003, thereby ensuring substantially smoother transitions in rapidly changing dynamic scenes. Ultimately, this framework provides a computationally efficient, highly robust perception solution for edge-deployed mobile robots, significantly enhancing operational safety and decision-making reliability in complex, human-centric environments.