毕业生知识库

In recent years, with the continuous development of computer vision and ethology, computational ethology, as a new interdisciplinary subject, has developed many automatic analysis systems of animal behavior. Because of their similar brain structure, macaques are used as an important experimental animal model for studying neuroscience and other fields. Therefore, the quantitative analysis system of macaque behavior in daily living environment can provide more sufficient and reliable experimental demonstration data for people to study the relationship between brain and behavior. Compared with other controlled environments, when collecting behavior information of macaques, animal behavior analysis in the daily living environment can avoid building new observation environments, thus reducing the observation cost. At the same time, the familiar environment can reduce the interference of non-natural factors caused by environmental changes, reduce the noise in the experimental data, and ensure the accuracy of the behavior analysis results. However, there is a relative lack of systems that can solve the problems faced by macaque behavior analysis in daily living environments. There are three main problems in the behavior analysis of rhesus monkeys in the daily living environment: occlusions, frequent changes in the environment and large spatial displacement when the monkey moves. In this paper, the above three problems are studied, and combined with the specific mainstream algorithm model to make targeted improvements. The main research contents and contributions of this paper are summarized as follows:

1. Realized stable and accurate trajectory tracking of macaques in daily living environment. Existing methods for track tracking of macaque monkeys are based on traditional image processing or deep learning. Both methods usually require a specific environment to reduce the interference of occlusion or environmental changes. A novel method MonkeyTrail is proposed in this paper. This method combines the frame difference method with the deep learn-based model (You Only Look Once, YOLO) to generate virtual empty background frequently, thus reducing the impact of environmental changes. After the empty background is generated, the method uses background subtraction to accurately obtain the differential pixels, i.e. the foreground area of the animal, so as to achieve tracking of the animal, where the differential pixels are not disturbed by the mesh occlusion. To test the performance of MonkeyTrail, a data set containing approximately 8 000 frames of video images (containing representations of macaques under various conditions) was used for comparison experiments. The results showed that the tracking accuracy and stability of MonkeyTrail exceeded that of two deep learn-based methods (YOLOv5 and Single-Shot MultiBox Detector), traditional frame difference, and naive background subtracting.

2. Realized stable and accurate behavior recognition of macaques in daily living environment. Among the current mainstream behavior recognition models, two-stream network and Transformer based network have better performance. However, there is interference information of metal reflection light in the daily living environment of macaques, which makes it difficult to get good results in the branch of optical flow in the two-stream network. At the same time, the spatial displacement of macaque movement leads to the loss of information interaction of moving parts in the Transformer method based on time series. Inspired by the basic logic of human observation and judgment of behavior, this paper built the Inter-Frame Action-area Attention (IFAA) module to focus on the movement area and movement trend of macaque monkeys. This module proposed the design of multi-layer activation of motion area with frame difference information, and enhanced the interaction of information between frames through frame difference to improve the spatial attention of motion parts and movement trends of macaques. At the same time, improving the attention of the moving area can also solve the problem that the accuracy of the model to recognize part of the behavior is decreased due to the occlusion. In order to train and test the model, by manually analyzing the daily behaviors of each macaque and considering the behavioral patterns concerned by the Institute of macaque behavior, this paper selected the most common 9 kinds of behaviors of macaque in the daily living environment, and finally produced a data set of about 5 hours of video format and original frame format. Finally, through ablation experiments, performance comparison with mainstream models and visualization analysis, this paper verifies that IFAA module can significantly improve the model's performance in the macaque behavior recognition task.

3. Set up an intelligent automatic analysis system for macaque motor behavior in daily living environment and verify its practicability. In order to facilitate researchers to train behavioral analysis models that meet individual needs, this paper builds an automated analysis system to modularize the process of video data quantitative analysis. Each module is a convenient application, including a data set processing module, an algorithm training module, and a prediction module. In order to meet the application requirements of low computing power platform, the system also adds distillation module to reduce the computing resources occupied by model prediction. In order to evaluate the application value of the system, this paper uses the data processing module of the system to produce a large number of video data of rhesus monkeys, and quantifies these data through the prediction module, including: (1) Video data of 2 macaques (5 days in each 1-year interval) were quantified into the amount of exercise and trajectory; (2) Video data of 7 macaques in a short period of time were quantified as the behavior time distribution; (3) Video data from 1 normal macaque and 2 Alzheimer's disease model monkeys (7 days before and after modeling) were quantified into the time distribution of specific behaviors. The prediction results of the above data prove that the system can provide normal monitoring for the daily behavior pattern of macaques and provide validation data for the behaviors of special concern in neuroscience experiments, which has good practicability.

In summary, effective algorithm design was carried out in this paper to solve the problems that occlusion, frequent environmental changes and individual motion displacement in the daily living environment of macaque monkeys lead to the difficulty of algorithm landing, and the trajectory tracking method and behavior recognition method were proposed which were superior to the mainstream methods in accuracy and stability. In addition, by building a behavior analysis system to modularize the overall data analysis process, and using the system to analyze a large number of daily behavior data and neuroscience experimental data, this paper verifies that the system has high practical value, so that researchers can easily conduct personalized data analysis by using low-cost hardware.