Helmet Pose Measurement Method Based on Feature Learning

In modern aviation systems, the helmet sight (HMS) assists the pilot to cooperate with the warplane to quickly complete the target tracking and aiming, and the helmet pose measurement is the core technology of the helmet sight. Since the helmet pose measurement technology involved in aviation safety, this technology is kept confidential at home and abroad, and there is a lack of open source helmet data set and high-precision helmet posture measurement algorithm. For this purpose, this paper produced the first helmet dataset, Helment 8K, and proposed the CSPDarkNet-Ht helmet pose measurement algorithm based on feature learning. The target is to form the helmet dataset image on the helmet model according to the unique feature points of the helmet model design. Helment 8K data set is made by building a helmet model with 3dmax and then rendering 8282 helmet images at different angles with targets. The CSPDarkNet-Ht algorithm divides the yaw angle and pitch angle of the helmet pose angle into two sub-tasks for classification and regression. Firstly, The algorithm uses CSPDarkNet to extract the low, medium and high semantic information of the target features for the backbone network. Then, different semantic feature layers are classified from different angles and regression of the highest semantic feature layer is performed to obtain the predicted pose angles. Finally, the loss function was optimized to include the cross-entropy loss of the five coarse and fine categories as the total loss. Moreover, the average absolute error of the angular prediction on the Helment 8K dataset is 0.5, and the prediction accuracy is 91.09% within 1 degree error, which is consistent with the accurate measurement of the helmet pose.