Lightweight design and deployment of object detection model for low-light scenes

Low-light scenes are common in fields such as autonomous driving, which tests the robustness of intelligent systems. This paper tests the performance of object detection in low-light scenes based on the YOLOv8n model. In order to solve the constraint of limited computing power of embedded devices, the YOLOv8n model is adjusted through model pruning and knowledge distillation. The complexity of the model is reduced through these lightweight operations. While ensuring the detection performance as much as possible, the multiply-accumulate operations (MACs) and parameters of the model decreased by about 40% and 27% respectively. The proposed model is evaluated on the Exclusively Dark (ExDARK) dataset and achieves a mean average precision (mAP) value of 0.672. In addition, the proposed model is migrated to the embedded platform NVIDIA Jetson Xavier NX. The deployment process is optimized by multi-process scheduling and multi-thread scheduling. Experiments are also conducted on real data, and the results show that the proposed model can provide a feasible solution for vision-based night time autonomous driving.