A Transformer-Based Approach to Efficient Dashcam Gps and Timestamp Extraction

• DOI: 10.2139/ssrn.5092224
• Publication Date: 2025-01-10T16:41:12Z
• Authors: Selvamani Chandra...; K. I. Ramachandran; Adarsh S; Binoy B Nair; Aravind J

Dashboard cameras are becoming increasingly prevalent in vehicles, leading to a sig-nificant demand for reliable methods extract essential metadata, such as timestamps,geolocation, and speed, from recorded footage. This metadata is contextual-izing events captured by these cameras, facilitating tasks accident reconstruction,security monitoring, forensic analysis. However, many modern low-cost dashboardcameras overlay text directly onto video rather than logging separately. Thisstudy aims develop solution that independent of specific devices extractingmetadata dashboard camera Optical Character Recognition (OCR) technol-ogy presents promising avenue due its hardware-agnostic capabilities. The researchinvestigates the application pre-trained OCR models, including Tesseract, KerasOCR,and EasyOCR, overlaid data images videos dashboardcameras, while also assessing limitations potential failure scenarios mod-els. findings indicate combination traditional algorithms prepro-cessing techniques achieves an average Rate (CRR) 50.6% anda Error (CER) 49.4%. To address shortcomings OCRmethods, novel Transformer-based (TrOCR) approachis proposed. Extensive training validation TrOCR model, utilizing mixeddataset real synthetic footage, significantly enhance character recognition accu-racy 84%, with corresponding reduction error rate 16%. Furthermore,the incorporation post-processing results exceptional accuracy 97%and negligible only 0.08%.