Maritime Autonomous Navigation

The maritime perception system employs a dual paradigm approach, using different detection methods for daytime and nighttime operation:

Nighttime Detection System (starting point)

The Light Spot Detection and Tracking (LSDT) system implements a sophisticated pipeline for identifying and tracking ship navigation lights at night.

The idea was to use the lights from the vessels to detect, track, and predict their position. Following international maritime regulations, the lights must be visible and the colors and their behavior indicate the type of vessel. Indeed, from the abstract: "the light spots in the video images are detected through LOG and invalid spots are filtered by the gray threshold. Multiple targets are subsequently tracked by Kalman filtering and light spots are marked to determine properties in order to add and delete spots."

The following outlines the method we started with. It is from an academic paper and allowed us to quickly get something working during occlusion. It was useful for the data association problem as well.

1. Light Spot Detection

Detection Pipeline (Validated on Yangtze River Dataset):

2. Multi-Target Tracking

# Light Spot Detection and Tracking System
LSDT_System {
    ImageProcessing [OpenCV]
    │── PreProcessing
    │   ├── Grayscale Conversion
    │   ├── Gaussian_Filter(σ=1.5)
    │   └── CLAHE_Enhancement
    │── SpotDetection
    │   ├── LOG_Filter(size=9x9)
    │   │   ├── Gaussian_Smoothing
    │   │   └── Laplacian_Operator
    │   ├── ZeroCrossing_Detection
    │   └── Threshold_Application
    └── SpotTracking
        ├── KalmanFilter
        │   ├── State_Vector[x,y,dx,dy]
        │   ├── Prediction_Step
        │   └── Update_Step
        └── DataAssociation
            ├── Hungarian_Algorithm
            ├── Gating(radius=30px)
            └── TrackManagement
                ├── Score_System
                ├── Track_Confirmation(n=5)
                └── Track_Termination(n=10)

Daytime Detection System

The daytime system leverages traditional RGB cameras with a specialized neural network trained on maritime scenes. We built a custom dataset of over 50,000 labeled maritime images to ensure accurate detection in various conditions and environments.

# Object Detection and Depth Estimation System
DaytimeVision {
    ObjectDetection [YOLO]
    │── Models
    │   ├── YOLOv3_MS_COCO
    │   ├── YOLOv4_Road_Obstacle
    │   └── YOLOv4_License_Plate
    │── BoundingBoxProcessing
    │   └── Centroid_Calculation
    └── DepthEstimation
        ├── MiDaS_MDE
        │   ├── Depth_Map_Generation
        │   └── Normalization
        └── Distance_Calculation
            ├── Thin_Lens_Model
            └── Relative_Depth_Mapping
}

The perception system was deployed on custom hardware designed for the maritime environment.

The collision avoidance system was designed to prevent near-miss incidents between vessels. It integrates advanced algorithms and data analysis to provide real-time recommendations for safe navigation.

The collision avoidance system was tested against 250+ historical near-miss incidents from the Northern European shipping lanes, demonstrating a 98.2% success rate in providing correct avoidance recommendations that complied with maritime regulations.

• Dynamic scenario testing for MDE approach
• LIDAR comparison and validation
• 3D object detection integration
• Country-specific vehicle datasets
• Integration of newer MDE models

For more information about the maritime navigation system, visit:

• Slalom Case Study