Design and Experiment of Real-Time Detection and Tracking System for Animal Behavior Based on Improved Retina Face Neural Network

Abstract

To solve traditional animal behavior recognition method defects such as non-intelligence, heavy workload, and poor generalization performance, improved RetinaFace neural network animal training model and NVIDA TensorRt were used to accelerate and construct the model engine. GPU asynchronous flow was used to process multiple frames of images and improve model-based reasoning speed. Finally, the real-time target detection and tracking system was built for the animal behavior. Experiment results showed that: The mAP value of the neural network model was 82.38%. Average time of forward reasoning model after improvement was 29ms and was increased by 3.58 times. The complete system can recognize and track the water maze environment in a paid and stable way and realize the real-time analysis on key behavior information of mice (such as trajectory, displacement, and exploration time).