基于数据增广的道路场景小目标检测

Small Object Detection in Road Scene Based on Data Augmentation

道路场景目标检测技术受限于数据集及目标检测算法,不同尺度目标检测精度差异显著,其中小目标检测性能较差。数据增广是解决该问题的主要手段,增加道路场景小目标训练样本,改变各尺度训练样本不均衡分布,提升其检测性能。针对等概率重采样存在局限性,提出随机概率重采样策略,增加了对小目标性能影响显著的训练图像。针对各尺度目标训练样本数量分布不均衡,提出自适应尺度均衡策略(Adaptive Scale matching Cutout,AdaSMC),缓解了大、中等目标被过度增广的问题。融合随机概率重采样和AdaSMC两种增广策略,提出应用于道路场景的融合增广算法。在Cityscapes数据集实验结果表明,该融合算法在保证实时性的前提下,APs提升1.9%,ARs提升1.7%。

The object detection technology of road scene is limited by dataset and object detection algorithm,and the detection accuracy of object at each scale is significantly different,among which the detection performance of small objects is poor.Data enhancement is the main method to solve this problem,which is to increase the training samples of small objects in road scenes,change the unbalanced distribution of training samples at various scales,and improve its detection performance.Aiming at the limitation of equal probability resampling,a random probability resampling strategy is proposed to add training images which have significant influence on the performance of small objects.Adaptive Scale matching Cutout(AdaSMC)is proposed to solve the problem that large and medium objects are over-augmented.Combining random probability resampling and AdaSMC augmentation strategies,a fusion augmentation algorithm for road scenarios is proposed.Experimental results of Cityscapes dataset show that APs and ARs can improve by 1.9%and 1.7%under the premise of real-time performance.