• Computer Vision techniques
• HAAR
• head pose
• HOG
• machine learning
• ORB
• pantograph detection
• template matching

Design and development of a vision system for detection and analysis of the pantograph head.
The proposed architecture detects, tracks, and analyzes the position of the pantograph head
and its geometrical structure to identify possible sources of defects that could be risky for train circulation.

The system records images, localizes relevant frames, compensates images color. Then, for the relevant images, the system detects the position and the spatial orientation of the pantograph head. Particular care has been posed to achieve precise shape and localization of the head. I developed a new processing pipeline that sequentially improves position accuracy. The pipeline exploits Computer Vision techniques such as template matching; machine learning (through HAAR cascade recognition and HOG objects detection; key features detection (SIFT and ORB algorithms); foreground extraction (GrabCut); color filtering with edge preservation; lines detection (Hough Transform).

Several parameters have optimized through relevant comparative analyses that address the accuracy of the final result and robustness of detection. Experimental results have been compared with a ground-truth composed of field data acquisition and result labeled by humans.
The achieved results demonstrated the capability of the system for being used as a decision support system in maintenance activities.