Application of a pupil tracking method based on Yolov5-Deeplabv3+ fusion network on a new BPPV nystagmus recorder

Abstract

Nystagmus information is an essential basis for diagnosing benign paroxysmal positional vertigo (BPPV), and nystagmus recorders are a crucial way to obtain nystagmus information. We designed a new wireless video nystagmus recorder, which uses the OV4689 sensor to collect the nystagmus video, encodes it on the RK3399 core control board, and sends it to the host computer through the 5GHz WiFi module. Compared with the current nystagmus recorder, it has the advantages of low price, convenient operation, and high frame rate. This paper uses a novel semantic segmentation network based on the Yolov5 target detection network and the improved Deplabv3+ segmentation module to perform real-time pupil segmentation on the collected video. Based on the segmentation results, we use ellipse fitting to extract the boundary and center of the pupil for drawing the pupil activity trajectory. Compared with the existing pupil tracking method, the method divides the pupil area faster and with higher precision, ensuring the pupil center positioning accuracy. The technique can effectively assist doctors in the diagnosis of BPPV.

Integrated neural network-based pupil tracking technology for wearable gaze tracking devices in flight training

“…The accuracy and stability of the network output results are verified by comparing them with the baseline network FPN, ViT, and the improved network DeepVOG model and HVit. Figures 14,15,and 16 show the pupil positioning accuracy of different networks after importing different data sets into FPN, ViT, DeepVOG network, HVit network, and our network respectively. We compare and verify the pupil positioning accuracy of each network by counting the error between the predicted value of the pupil center output by different networks and the actual value when it is less than that of different pixels.…”

“…The method proposed by Wei K achieves good results in pupil segmentation, but this method does acquisition not take into account the detection accuracy of pupils in dark light environments. To obtain the position information of the subject's eyes in different states, Newman J L [16] and others built a 2D convolutional neural network with the help of CAVA equipment. The researchers fused the head and eye movement data of patients when they experienced vertigo and realized vertigo detection based on the fusion of eye movement signals and head posture.…”