摘要
针对现有的CornerNet-Saccade算法在车辆检测任务中存在明显的误检和漏检现象,提出了一种改进的CornerNet-Saccade算法。首先通过加深堆叠沙漏网络结构,增强车辆高级特征提取能力;其次,增加更小尺度的attention maps以改善小目标车辆的检测能力,引入Dense Block模型和瓶颈残差单元降低堆叠沙漏网络参数的复杂度,保证层与层之间的最大信息流;最后,通过Sigmoid激活函数得到最终的检测结果。在KITTI数据库和自制数据库上对改进算法进行了仿真实验,平均精确率分别达92.56%和95.21%,检测速度分别达40 FPS和49FPS,同时在自制数据库上对原CornerNet-Saccade算法和改进算法进行了仿真实验,精确率和召回率相比原算法分别提高了3.8%和8.5%。结果表明:此改进的CornerNet-Saccade算法在车辆检测任务中具有明显优势。
In view of the obvious phenomenon of false detection and missing detection by using the existing CornerNet-Saccade algorithms,an improved CornerNet-Saccade algorithm was proposed in the vehicle detection task.First,the Stacking Hourglass Network structure was deepened to enhance the capability of vehicle advanced feature extraction.Then the smaller attention map was used to improve the detection capability of small target vehicles,the Dense Block model and Bottleneck Residual Unit were introduced to reduce the complexity of Stacked Hourglass Network parameters and ensure the maximum information flow between layers.Finally,the detection result was obtained by using the Sigmoid activation function.This improved algorithm achieves average precision of 92.56%and 95.21%and detection speed of 40 FPS and 49 FPS on the KITTI and homemade datasets,respectively.Meanwhile,the original CornerNet-Saccade algorithm and the improved algorithm were simulated on the homemade database,the precision and recall rate were improved by 3.8%and 8.5%compared with the original algorithm.The results showed that the improved CornerNet-Saccade algorithm can perform better in the vehicle detection task.
作者
梁礼明
熊文
蓝智敏
钱艳群
LIANG Liming;XIONG Wen;LAN Zhimin;QIAN Yanqun(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2021年第6期137-146,共10页
Journal of Chongqing University of Technology：Natural Science
基金
国家自然科学基金项目(51365017,61463018)
江西省自然科学基金面上项目(20192BAB205084)
江西省教育厅科学技术研究重点项目(GJJ170491)。
关键词
车辆检测
堆叠沙漏网络
注意力机制
锚点框
角点检测
vehicle detection
stacked hourglass networks
attention
anchorbox
corner detection
