基于改进SSD算法的车辆检测

SSD算法利用多尺度特征图进行分类和位置回归，检测小目标效果优于YOLO算法，但SSD算法在进行车辆检测时存在漏检问题。为此，提出一种改进SSD算法。为提取更多的车辆特征信息，设计改进Inception模块替代SSD网络中的Conv8、Conv9和Conv10层。将浅层特征的位置信息和深层特征的语义信息进行均衡化融合，构建多尺度特征融合均衡化网络，提高小目标车辆识别率。在特征提取层均引入SENet，对不同特征通道的重要性进行重标定以提高模型性能。实验结果表明，改进后SSD算法在自制的车辆数据集上平均精度为90.89%，检测速度达到59.42 frame/s，相比改进前的SSD算法，在精度和速度上分别提高2.65个百分点和17.41 frame/s，能够更快速、准确地对图像中的车辆进行识别和定位。

Vehicle Detection Based on Improved SSD Algorithm

The Single Shot Multibox Detector(SSD) algorithm uses multi-scale feature maps for classification and position regression.This algorithm displays better performance than the YOLO algorithm in detecting small targets, but frequently misses targets in vehicle detection.To address the problem, this paper proposes an improved SSD algorithm. In order to extract more vehicle feature information, an improved Inception module is given to replace the Conv8, Conv9 and Conv10 layers in the SSD network.At the same time, a network is designed to balance and fuse multi-scale features.The network can fuse the location information of shallow features and the semantic information of deep features, so the accuracy of small target detection can be increased.To further improve model performance, SENet is introduced into feature extraction layers to recalibrate the importance of different feature channels.The experimental results show that the improved SSD algorithm displays an average accuracy of 90.89% and a detection speed of 59.42 frame/s on the self-made vehicle dataset.Compared with the original SSD algorithm, the improved SSD algorithm increases the accuracy by 2.65 percentage points and speed by 17.41 frame/s.This algorithm can identify and locate the vehicles in images more quickly and accurately.