Image Tracking of Fire Extinguishing Jet Drop Point Based on Improved ENet and Robust Adaptive Cubature Kalman Filtering

Accurate image tracking of fire extinguishing jets is crucial to achieving automatic firefighting. However, inevitable noise interference occurs during image processing, adversely affecting precise tracking. In order to address this issue, a method for tracking the jet drop point (JDP) of a fire extinguishing jet is proposed based on an improved efficient neural network (ENet) and robust adaptive cubature Kalman filter (CKF). A novel JDP image state transition model is established to construct the state space equations and depict the motion state of the JDP in images. A two-stage method for recognizing JDP is proposed, which includes an improved ENet and a directional progressive curve search method to enhance the accuracy of observation. A CKF based on the Huber function is proposed to improve the adaptability and robustness of the image tracking method, which takes into account the advantages of $L1$ and $L2$ norms. The updated formulas for the state and covariance matrices are derived. Furthermore, the tracking method is improved by the Sage-Husa method, which considers the unknown distribution of noise. Experiments on actual firefighting platforms demonstrate that the proposed method exhibits robustness and adaptability compared to traditional CKF.