Effects of Hybridizing the U-Net Neural Network in Traffic Lane Detection Process

A vision-based lane detection system is difficult to implement in vehicles due to the unique characteristics of algorithms, neural network architectures, limitations, and strict hardware/software requirements. This paper proposes a new architecture, called hybrid U-Net, which represents a significant extension of the classical U-Net model, addressing its limitations in capturing global relationships, adaptability to scale variations, and sensitivity to geometric structures. An important aspect of this new architecture is the balance between performance and computational efficiency, maintaining reasonable values of performance parameters with a competitive inference time. The experimental results show that hybrid U-Net achieved superior scores in both mIoU and F1 score parameters. The most notable increase was observed in the (TuSimple+Carla) dataset, where mIoU increased from 0.56 to 0.60 (+7%), and F1 score from 0.69 to 0.74 (+7.2%). Similarly, in the Carla dataset (known for the complexity of the generated images), a substantial improvement was recorded, with an increase of +8.0% in mIoU and +5.7% in F1 score, showing better adaptability of the model to geometric structures in complex images. However, these improvements come at a considerable computational cost, as the inference time of the hybrid U-Net model is approximately 0.04 s per image (compared to 0.016 s per image for U-Net), resulting in a decrease in FPS from 60 to approximately 23. Future research should investigate model compression techniques, such as knowledge distillation, pruning, or quantization, to improve efficiency without sacrificing accuracy.