Vehicle Re-Identification (V-ReID) menjadi tantangan utama dalam sistem pengawasan lalu lintas akibat variasi tampilan kendaraan yang signifikan pada skenario kamera tidak tumpang tindih. Kesenjangan tampilan yang kompleks antar kamera membuat fitur visual kendaraan sulit dikenali secara konsisten. Penelitian ini mengembangkan model V-ReID berbasis handcrafted features dengan mengusulkan algoritma Oriented FAST and Rotated BRIEF (ORB) yang telah dioptimalkan guna meningkatkan akurasi dan efisiensi.
Kontribusi utama dalam penelitian ini terletak pada perancangan Improved ORB, yang mencakup: (1) pembagian citra ke dalam beberapa wilayah guna memastikan distribusi keypoint yang lebih merata; (2) penerapan thresholding adaptif berbasis statistik intensitas untuk menyaring fitur-fitur yang lebih relevan; dan (3) integrasi metode Difference of Gaussian (DoG) dari piramida Gaussian milik SIFT untuk meningkatkan ketahanan terhadap perubahan skala dan rotasi. Selain itu, digunakan seleksi keyframe berbasis Mutual Information Entropy untuk mengurangi redundansi dan mempercepat pemrosesan video.
Evaluasi pada dataset VeRI-FMIPA dan VeRI-776 menunjukkan bahwa Improved ORB secara signifikan melampaui ORB standar, dengan Top-1 accuracy sebesar 65,56?n mAP 65,97% pada VeRI-FMIPA, serta 62,12?n 63,88% pada VeRI-776. Hasil ini membuktikan bahwa pendekatan local feature descriptor yang dioptimalkan tetap relevan dan kompetitif dalam kondisi keterbatasan data latih dan kebutuhan komputasi real-time, serta berpotensi dikembangkan lebih lanjut untuk sistem pengawasan lalu lintas cerdas.
Vehicle Re-Identification (V-ReID) has emerged as a major challenge in traffic surveillance systems due to the significant variation in vehicle appearances across non-overlapping camera scenarios. The complex visual gap between camera views makes it difficult to consistently recognize vehicle features. This study proposes a V-ReID model based on handcrafted features by introducing an optimized Oriented FAST and Rotated BRIEF (ORB) algorithm to enhance accuracy and efficiency.
The main contribution of this study lies in the design of the Improved ORB, which includes: (1) dividing the image into multiple regions to ensure a more even distribution of keypoints; (2) applying adaptive thresholding based on intensity statistics to filter out more relevant features; and (3) integrating the Difference of Gaussian (DoG) from the SIFT Gaussian pyramid to enhance robustness against scale and rotation variations. In addition, a keyframe selection method based on Mutual Information Entropy is employed to reduce redundancy and accelerate video processing.
Evaluations on the VeRI-FMIPA and VeRI-776 datasets demonstrate that the Improved ORB significantly outperforms the standard ORB, achieving a Top-1 accuracy of 65.56% and mAP of 65.97% on VeRI-FMIPA, and 62.12% and 63.88% respectively on VeRI-776. These results confirm that the optimized local feature descriptor approach remains relevant and competitive under limited training data and real-time processing constraints, and holds strong potential for further development in intelligent traffic surveillance systems.
Kata Kunci : Vehicle Re-Identification (V-ReID), Deskriptor Fitur Lokal, ORB, Piramida SIFT Gaussian, Pengawasan Lalu Lintas
