Abstract:

In order to improve the detection performance of one-stage object detection algorithm for small objects and overlapping objects, and make it applicable to automatic driving scenarios, a deep feature fusion algorithm DFSSD based on SSD（Single Shot Multibox Detector） is proposed. DFSSD mainly improves the SSD algorithm from two aspects：on one hand, an efficient feature fusion method is adopted to enhance the feature expression ability of the model and the detection ability for difficult small objects without introducing extra parameters and excessive computation. On the other hand, a training method with noise is introduced, that is, during the training, the unmarked hard positive samples（undistinguishable positive examples） are randomly added to the training, so as to improve the anti-interference ability of the algorithm to complex background and reduce false detection rate of small difficult objects. On the PASCAL VOC2007 test set, compared with SSD300, DFSSD300’s mAP（mean Average Precision） increases by 3.7 percentage points, and on the KITTI dataset, the AP（Average Precision）  of  difficult objects in car class increases by 5 percentage points. At the same time, it has the same detection speed as SSD300.

Key words: deep feature fusion, small object detection, deep learning, automatic driving scene

摘要：

为了提高单阶段目标检测算法对小目标和重叠目标的检测性能，使其能够应用到自动驾驶场景中，提出一种基于SSD（Single Shot Multibox Detector）的深度特征融合算法DFSSD（Deep Fusion based Single Shot Multibox Detector）。DFSSD主要从两个角度对SSD算法进行改进：一方面提出一种高效的特征融合方式，在不引入大量参数和过多计算量的情况下，增强了模型的特征表达能力和对困难小目标的检测能力；另一方面引入一种带噪声的训练方式，即在训练时，随机地将样本中未标记的困难正例目标（不易分辨的正例目标）加入训练，以提高算法对复杂背景的抗干扰能力，降低对困难小目标的误检率。在PASCAL VOC2007测试集上，DFSSD300比SSD300的mAP（mean Average Precision）提升了3.7个百分点，在KITTI数据集上，Car类困难目标的AP（Average Precision）值提升了5个百分点，同时与SSD300具有相当的检测速率。

关键词: 深度特征融合, 小目标检测, 深度学习, 自动驾驶场景