基于K-SSD-F的东亚飞蝗视频检测与计数方法

Video Detection and Counting Method of East Asian Migratory Locusts Based on K-SSD-F

针对目前国内蝗虫监测主要以人工监测为主、监测效率低且计数不准确的问题,以5龄东亚飞蝗为实验对象,提出了一种蝗虫视频计数方法K-SSD-F算法。该方法可以实时、连续、自动监测蝗虫的数量。首先利用背景分离法中的KNN算法提取视频前后帧的时空特征;然后通过标注好的数据训练SSD模型,并对视频进行检测,提取视频的静态特征,二者结合以提高计数准确率;最后利用补帧算法识别因姿态变化导致的漏计数的帧。实验结果表明,蝗虫识别准确率为97%,召回率为89%,平均检测精度(mAP)为88.94%,F1值为92.82%,且检测速度达到了19.78 f/s。本文方法具有较好的鲁棒性,可以实现蝗虫的实时和自动计数,其精度优于其他模型,也可为其他种类的昆虫自动识别计数提供理论基础。

At present,domestic locust monitoring is mainly based on manual monitoring,with low monitoring efficiency and inaccurate counting.In response to the above problems,the K-SSD-F algorithm,a video counting method of locusts,was proposed with the 5 th instar migratory locust as the experimental object.This method can monitor the number of locusts in real time,continuously and automatically.Firstly,the KNN algorithm in the background separation method was used to extract the spatiotemporal features of the frames before and after the video;then the SSD model was trained through the labeled data,the video was detected,and the static features of the video were extracted,and the two were combined to improve the counting accuracy;finally,the frame compensation algorithm was used to recognize missing frames due to posture changes.The experimental results showed that the precision of locust identification was 97%,the recall rate was 89%,the average detection accuracy(mAP)was 88.94%,the F1 value was 92.82%,and the detection speed reached 19.78 f/s.The proposed method had good robustness,which can realize real-time and automatic counting of locusts,its accuracy was better than that of other models,and it can also provide a theoretical basis for automatic identification and counting of other kinds of insects.