Deep Learning—Based Image Matching Algorithm with Improved Detection and Tracking Performance

Template matching is an approach to locating parts of an image that are similar to a template. It is often employed in computer vision applications such as object detection, pattern recognition, image registration, video surveillance and security, medical treatments, robotics, and smart automobiles. Pattern matching is also essential in object tracking, which involves identifying and following the object(s) across video frames while adapting to changes in appearance, occlusions, and lighting. Although many tracking algorithms have been developed, the problem of object matching (re-identification) remains valid and requires further research. Notably, matching tasks in which the object class is predefined can be performed using Siamese neural networks. Our main goal is to propose a deep learning-based matching algorithm that can be seamlessly integrated into large computer vision pipelines on embedded devices, ensuring optimal performance in real-time edge applications. We compose a dataset from publicly available and in-house data to test the proposed method. We employ conventional matching algorithms and methods based on convolutional neural networks (CNNs) and Siamese neural networks to compose feature vectors (embedding) representing the image. We measure the similarity and evaluate the performance of the proposed matching algorithm using different distance metrics on image embeddings. The proposed Siamese-based matching method shows higher performance than the CNN-based one. The proposed method was incorporated into a computer vision pipeline and evaluated on NVIDIA Jetson family embedded devices in real-time processing scenarios, achieving an average inference performance of 40 Hz. This matching model can be utilized in real-time video analytics to enhance the reliability and robustness of the analysis.