Channel Attention Based Knowledge Distillation for Segmentation of Country Club Driving Environments

Recent advancements in deep learning based visual recognition methods have gained significant attention in a country club driving environment, particularly in semantic segmentation. Autonomous driving systems require semantic segmentation methods to understand the type, location, and shape of objects and for accurate recognition of the driving environment. However, state-of-the-art methods are computationally resource intensive and lightweight segmentation methods have poor performance, limiting their applicability to embedded systems. This paper proposes an attention module based knowledge distillation method, lightweight channel attention distillation (LCAD) to improve semantic segmentation accuracy in lightweight embedded systems. Moreover, we propose a lightweight channel attention module (LCAM), which uses global average pooling and 1D convolution to reduce the number of parameters efficiently. Experimental results, obtained using the custom dataset collected from a real-world golf course environment, demonstrate that our method achieves improved segmentation accuracy with reduced model complexity compared to existing knowledge distillation methods, and the efficiency of lightweight visual recognition models for autonomous golf cart navigation.