摘要
安防场景的行人目标检测不仅需要识别并标定行人位置,还需要对人脸进行检测与提取,支持人脸比对等下游任务。当前目标检测算法在实际应用中存在运行速度慢、边界框标定不准确、小目标检测效果不佳等问题,文中提出一种基于SSD检测器的安防场景下的行人目标检测算法。针对检测器因为分类与定位子任务强耦合造成边界框标定不准确的问题,采用一种解耦的“检测头”保证检测器定位精度,并通过在不同分支上引入特征增强模块提取适应不同子任务的特征;采用一种任务耦合的损失函数来提升训练效果;针对运行速度慢,采用轻量化网络作为主干网络,结合TensorRT量化模型提升算法在嵌入式平台上的运行速度。通过在NVIDIA TX2嵌入式深度学习平台进行实验,单帧图像检测时间为23.8 ms,平均帧率约为42 f/s,算法具备优秀的实时性与准确性。
Pedestrian object detection for security scenario requires not only identifying and calibrating pedestrian locations,but also detecting and extracting faces to support downstream tasks such as face matching.The current object detection algorithm has some deficiencies in practical application,such as slow running speed,inaccurate bounding box calibration and poor effect of small object detection.Therefore,an SSD⁃based(single shot multibox detector based)pedestrian object detection algorithm for security scenario is proposed.To address inaccurate bounding box calibration due to the strong coupling between classification and subtask localization,a decoupled detection head is used to ensure localization accuracy,and features adapted to different subtasks is extracted by introducing feature enhancement modules on different branches.A task⁃coupled loss function is adopted to improve the training effect.To address the running speed,a lightweight network is adopted as the backbone to improve the running speed of the algorithm on the embedded platform in combination with the TensorRT quantization model.Experiments on NVIDIA TX2 embedded deep learning platform show that the single⁃frame image detection time of the proposed algorithm is 23.8 ms,and its average frame rate is about 42 f/s.Therefore,the algorithm is of excellent real⁃time performance and accuracy.
作者
周文凯
李佳怡
周建华
樊中财
ZHOU Wenkai;LI Jiayi;ZHOU Jianhua;FAN Zhongcai(Zhejiang Dahua Technology Co.,Ltd.,Hangzhou 310056,China)
出处
《现代电子技术》
2023年第21期59-63,共5页
Modern Electronics Technique
