Evaluation of object detection algorithms for ship detection in the visible spectrum

Abstract

The research described here examined computer vision algorithms for suitability to aid or replace the current methods of ship detection and tracking from a photonics mast. Evaluation was conducted on three object detection methods: a bag of words (BOW) robust multi-class classification method; a histogram of oriented gradient (HOG) method, originally used for pedestrian tracking; and a deformable parts model (DPM) that was originally designed for pose recognition that has been successful in multi-class classification. A fourth method that combines the HOG and BOW was created and successfully reduced false positive detections while maintaining a high recall rate. The object detection methods were evaluated through a search theory model to frame evaluation for operational ship detection. Each object detection method was optimized following a design of experiments approach utilizing a cluster computer. The BOW method had the highest recall for ships 25 pixels and smaller, while the HOG method was the fastest of all methods when implemented on a graphical processing unit. The DPM method had the highest average recall for ships greater than 25 pixels but the lowest recall for smaller ships. Finally, the hybrid HOG and BOW method had the highest mean recall and lowest mean false positive rate over all ship sizes.