REAL-TIME MONOCULAR MARITIME OBJECT DETECTION, TRACKING, AND DISTANCE–VELOCITY ESTIMATION UNDER OPEN-WATER CONDITIONSPages 74-80 Abstract
Maritime navigation systems require continuous perception of surrounding vessels and floating objects to support collision awareness and proximity assessment. Image-space detection metrics alone do not provide physically interpretable range or motion information. In the absence of stereo or active depth sensors, monocular vision must rely on geometric projection constraints to infer metric distance. Sensitivity of such estimation increases as object pixel scale decreases, particularly in open-water scenes where target size varies significantly with range. This paper fills the gap by implementing and evaluating of a real-time monocular maritime perception pipeline integrating object detection, multi-object tracking, and geometric distance–velocity estimation. Detection is performed using a YOLO-based one-stage architecture at 640×640 resolution. Tracking employs a constant-velocity state model with IoU-based association. Distance is computed using a calibrated pinhole projection model with class-level height priors. Radial velocity is derived through temporal differencing of estimated range. Experiments were conducted using Ultralytics (PyTorch backend) on NVIDIA A100-SXM4-40GB hardware. Detection achieved average precision of 0.949 and F1 score of 0.92.
Keywords:
Monocular maritime perception,
YOLO-based object detection,
Multi-object tracking (SORT),
Pinhole projection ranging,
Real-time distance–velocity estimation.
|
Ye Si Thu Aung, 
PDF 