SiamMCE: An Efficient Siamese Network for Real-Time Object Tracking With Dual-Correlation Strategy and Dynamic Template Updating

Abstract

Real-time object tracking in dynamic environments poses significant challenges in balancing computational efficiency with robust performance under complex scenarios such as occlusion and illumination changes. This paper presents SiamMCE (Siamese MobileNet-CE) tracker, an optimized Siamese network variant that integrates three key innovations to address these challenges. First, we design a lightweight backbone network MobileNet-CE which is based on MobileNetV3-Small through strategic integration of Convolutional Block Attention Module (CBAM) and Efficient Channel Attention Module (ECAM), reducing parameters by 18% while enhancing feature discriminability. Second, we propose a Dual-Correlation Strategy combining Pixel-wise Correlation (PWC) and Channel-wise Correlation (CWC) operations to improve localization precision through complementary spatial-channel feature fusion. Third, a Dynamic Template Adaptation mechanism leverages response map analysis via UpdateNet to enable online template refinement, mitigating drift accumulation during long-term tracking. Extensive experiments on benchmarks (OTB-2015, VOT2018, UAV123) demonstrate that SiamMCE achieves robust performance across mainstream tracking tasks, balancing competitive accuracy with real-time operation on embedded platforms. This capability enables new applications in dynamic environments, such as UAV-based detection and mobile surveillance, where sustained reliability is critical.