Posture Estimation Research On Grooming Behavior Of Adult Bactrocera Minax Via Limb Key Points Tracking

Bactrocera minax is the main pest of citrus,which seriously affects fruit yield and human health for a long time.It has been confirmed that inhibition of grooming behavior increases mortality in insect-pathogen bioassay.So statistical analysis of Bactrocera grooming behavior is important for pest control and human health.The traditional tracking method of insect limb movement belongs to invasive detection,for instance,the use of artificial pheromone may affect the behavior of Bactrocera minax,thus directly affect the accuracy and reliability of experimental results.Traditional grooming behavior researches mainly relies on manual work for behavior analysis and statistics.Researchers need to play the video frame by frame and record the time interval of each grooming behavior manually,which is time-consuming,laborious,and inaccurate.So,the advantages of automated analysis are obvious.Based on Deep Lab Cut,this paper proposes a non-invasive and effective intelligent method to track the key points of Bactrocera minax,to detect and analyze its grooming behavior.The main research contents are follows:(1)Study a variety of data enhancement methods and compare their advantages and disadvantages.According to the strong sensitivity of Deep Lab Cut algorithm to the body structure and pattern of Bactrocera minax,conventional rotation and mirroring strategies were selected to expand the original Bactrocera minaximage data.(2)By comparing and analyzing common feature extraction networks,testing network performance under the same data set,analyzing advantages and disadvantages,and finally selecting Res Net-50 as the algorithm backbone network.Compared with common multi-target tracking algorithms,Deep Lab Cut is selected as the key points tracking algorithm model of Bactrocera minax to capture the movement trajectory of fourteen key limb points,such as forefoot,middle foot,hind foot,antenna,head,abdomen and wings,and the coordinate position deviation of outliers in the trajectory is corrected by the mean function.(3)The grooming behavior classification model was established according to the movement track of each key point,and the behavior sequence was counted.Finally,the self-built data smoothing algorithm was used to eliminate and fill the error detection frame,and the statistical grooming behavior sequence was smoothed and optimized.Using the method proposed in this paper,a total of 94538 frames of grooming behavior image data were analyzed and fourteen limb key points were tracked.The overall tracking accuracy was as high as 96.7%.The overall recognition accuracy reached94.32% for the seven kinds of grooming behaviors: forefoot grooming,hind foot grooming,Forefoot and midfoot reciprocal grooming,midfoot and hindfoot reciprocal grooming,wings grooming,abdomen grooming and antennae grooming.The experimental results show that the automatic non-invasive method designed in this paper can track the many limbs key points of Bactrocera minax with high accuracy,ensure the accuracy of insect behavior recognition and analysis,which greatly reduces the manual observation time,and provides a new method for key points tracking and behavior recognition of insects.