Standalone package delivery verification using computer vision and image metadata extraction

Chow, Yong Xiang (2026) Standalone package delivery verification using computer vision and image metadata extraction. Final Year Project, UTAR.

PDF Download (41Mb)

Abstract

Last-mile delivery is widely recognized as the most complex and expensive stage of the e-commerce supply chain, often characterized by high fragmentation and limited observability [1]. Challenges such as "porch piracy", the theft of unattended packages, and the lack of structured Proof-of-Delivery documentation lead to significant operational inefficiencies, including payment delays and unresolved customer disputes [2], [3]. This project addresses these vulnerabilities by introducing a Zero-Trust Visual Verification Pipeline delivered through a standalone React Native mobile application. The system utilizes Exchangeable Image File Format metadata as a core novelty to enhance delivery verification. By extracting and validating high-precision geolocation and timestamp data at the point of capture, the application ensures that delivery evidence reaches a forensic-grade standard of precision [4]. This "Gated Capture" mechanism is further reinforced by a Siamese Neural Network, which provides temporal image validation by measuring the similarity between delivery photos to verify parcel identity across different stages [5], [6]. To maintain operational continuity in diverse environments, the architecture includes an adaptive accuracy algorithm for GPS-denied scenarios and a building detection model optimized for edge deployment. Developed within the Expo Application Services ecosystem and supported by a Supabase cloud backend, this system provides a secure, tamper-resistant framework for logistics tracking. By binding parcel unique identifiers with verified environmental context, the proposed solution transforms delivery records into legally defensible assets, ultimately strengthening transparency and reducing fraudulent claims in last-mile logistics [7].

Actions (login required)