While the convolutional layer deepens during the feature extraction process in deep learning networks, the performance of the object detection decreases associated with the gradual loss of feature integrity. In this paper, the convolutional feature frequency adaptive fusion object detection network is proposed to effectively compensate for the missing frequency information in the convolutional feature propagation. Two branches are used for high- and low-frequency-domain channel information to maintain the stability of feature delivery. The adaptive feature fusion network complements the advantages of missing high-frequency features, enhances the feature extraction integrity of convolutional neural networks, and improves network detection performance. The simulation tests showed that this algorithm’s detection results are significantly enhanced on blurred objects, overlapping objects, and objects with low contrast between the object and background. The detection results on the Common Objects in Context dataset was more than 1% higher than the CornerNet algorithm. Thus, the proposed algorithm performs well for detecting pedestrians, vehicles, and other objects. Consequently, this algorithm is suitable for application in autonomous vehicle systems and smart robots.

虽然在深度学习网络的特征提取过程中卷积层会加深，但目标检测的性能会随着特征完整性的逐渐丧失而下降。本文提出了卷积特征频率自适应融合目标检测网络，以有效补偿卷积特征传播中丢失的频率信息。两个分支分别用于高、低频域信道信息，以保持特征传递的稳定性。自适应特征融合网络弥补了高频特征缺失的优点，增强了卷积神经网络的特征提取完整性，提高了网络检测性能。仿真测试表明，该算法对模糊物体、重叠物体、物体与背景对比度低的物体的检测效果显着增强。Common Objects in Context数据集的检测结果比CornerNet算法高1%以上。因此，所提出的算法在检测行人、车辆和其他物体方面表现良好。因此，该算法适用于自动驾驶汽车系统和智能机器人。