Research And Application Of Video Object Detection For Dairy Goat Based On Temporal Information

Using intelligent video surveillance technology to monitor dairy goat farms helps to obtain flock dynamics and abnormal behaviors in a timely manner,improve the management efficiency of goat farms and reduce farming costs.Video object detection is an important part of intelligent monitoring technology,and its detection results directly affect the effectiveness of subsequent object tracking and object behavior analysis.In this paper,the video object detection of dairy goats based on the improved Faster R-CNN algorithm is implemented based on the dairy goat monitoring video from the animal husbandry teaching experiment base of Northwest Sci-Tech University of Agriculture and Forestry.The main research contents and conclusions of this paper are as follows:(1)Construction of dairy goat video object detection dataset.To address the problem of insufficient dairy goat video object detection dataset,firstly,remote cameras are installed in dairy goat farms to obtain the original dairy goat surveillance video;secondly,FFmpeg is used to frame the surveillance video and Label Img open source tool is used to label the frame images;finally,the files are organized according to ILSVRC-VID dataset format and the final dairy goat dataset is named GOAT-Dataset,which is used for the training of subsequent algorithm models.(2)Feature-weighted video object detection of dairy goat.A feature-weighted dairy goat video object detection model based on SF-Faster R-CNN is proposed to address the problem of inadequate utilization of video frame feature information by convolutional neural networks.SF-Faster R-CNN embeds an improved SE channel attention module based on Faster R-CNN and Res Net-50 backbone network for improving the learning weight of convolutional neural network for effective features;the multi-scale feature fusion FPN module is introduced to make the extracted features more richly expressive.It is experimentally demonstrated that the SF-Faster R-CNN has better detection performance and the detection accuracy reaches 71.54%on GOAT-Dataset dairy goat dataset.(3)Dairy goat video object detection based on aggregated temporal information.A dairy goat video object detection model based on aggregated temporal information is proposed for the problem that the SF-Faster R-CNN model does not take full advantage of the high similarity of the same object features between adjacent frames.This model divides video frames into key and non-key,uses feature extraction network for feature extraction on key frames only,and enhances key frame features by recursive feature aggregation,and for nonkey frame features,it is obtained by feature propagation using a lightweight optical flow network;then the Temporal ROI Align operator is introduced,which uses feature similarity in time series.This operator extracts the features of the current frame candidate from the feature mapping of other frames using feature similarity on the time series.It is experimentally demonstrated that the dairy goat video object detection model based on aggregated temporal information has high accuracy and fast recognition speed on GOAT-Dataset dataset.Finally,the dairy goat video object detection platform based on temporal information is designed for online detection of dairy goat surveillance videos.