In low-light environments, detection tasks are challenging due to complex lighting conditions and poor image quality with low signal-to-noise ratios from vehicle-mounted cameras, leading to poor perception performance. The use of general object detection methods results in low accuracy. To address these issues, this paper proposes a nighttime algorithm, P-Net, based on the lightweight network MobileNetV3. The algorithm features a parallel feature extractor, which fuses deep feature information between feature layers of different scales to enhance image quality. Additionally, a road dataset called BDD-dark, containing 5200 paired daytime-nighttime samples, was created for this study. On this dataset, the object detection task accuracy was improved by 1.5%, and the algorithm’s detection time per frame is only 21.97 ms, meeting real-time detection requirements.