Research On Deep-learning Object Tracking Technology For Optogenetics Irradiation System

Optogenetics technology is an important tool in neuroscience research.However,the experimental systems based on wired optical fiber impose great constraints on animals,and they are difficult to apply to animal behavioral researches.In this thesis,to realize unconstrained light stimulation to animals,the object tracking algorithm based on deep learning is studied and designed for the animal locating requirement in the optogenetics laser irradiation system.Aiming at the problem of missing features caused by the changeable posture and strong randomness of movement of experimental animals,an object tracking network based on generative adversarial learning is proposed,and it is called ATNet.The residual attention module is introduced into the object tracking network to improve the feature extraction process of the target,which highlights the features of the target area and suppresses the features of the background area in the experimental scene.And the network is combined with the generative adversarial learning method,where the mask generation network learns how to identify the key features of the target from image features and attentional information,so as to enhance the positive samples in training to improve the ability of the tracking network to identify samples with missing features.Semantic regions are selected as negative samples according to attentional information,which reduces the drift of the tracker and improves the locating precision.To solve the problem of losing targets caused by animals leaving the field of view,a long-term object tracking algorithm based on anomaly detection and its startup strategy is proposed.Abnormal tracking results are detected according to the cosine similarities between the image samples,and a dynamic threshold reflecting the violent degree of the target change is used to adapt to different targets.The sample rehearsal method reduces the tracker’s forgetting of the target features,which improves the ability of the tracker to correctly recognize the target after the loss occurs.When the target is lost,the candidate regions are proposed according to the contour features of the local regions to accelerate re-detection and improve the locating effect during tracking.This thesis carries out tests on the public tracking datasets and the optogenetics experimental mouse image datasets.The experimental results show that the proposed object tracking network based on adversarial learning has a good tracking precision,the average tracking error is reduced from 16.96 pixels to 8.48 pixels,and the percentage of time that its precision reaches the standard is increased from 36.3% to 57.9%compared with other tracking methods.In the scene where the target is out of bounds,by introducing the long-term tracking algorithm based on anomaly detection,the average tracking error is reduced from 13.24 pixels to 6.78 pixels,and the percentage of time that its precision reaches the standard is increased from 48.4% to 87.4%.The experimental results show that the proposed object tracking algorithms could meet the requirements of the optogenetics laser irradiation system for locating experimental animals and is conducive to the implementation of the wireless optogenetics experiments.