Deep learning-based detectors are essential in autonomous driving perception systems, allowing vehicles to identify objects and their locations while providing critical sensory information for decision-making. However, small objects like traffic cones and pedestrians often lead to higher false detection rates. To address this, the industry typically expands the training dataset, improving accuracy but incurring high annotation costs. Therefore, exploring cost-effective methods to reduce false detections is vital. We propose an optimization strategy using auxiliary segmentation networks, adding an auxiliary branch to the existing network. This approach leverages existing annotation data to generate segmentation mask ground truth, offering stronger supervision for detecting specific objects. Model using this strategy achieved a 73.27% reduction in false detections on the cone negative sample evaluation set.