Multicamera multitarget tracking based on lightweight neural network

Multi-Camera Multi-Target Tracking (MCMT) is critical for environmental perception in intelligent transportation and autonomous driving. However, existing BEV-based MCMT models face critical deployment challenges: (1) heavy backbones like ResNet result in excessive parameters and computation costs that exceed limited in-vehicle hardware capabilities; and (2) robustness verification for single-view loss (e.g., camera occlusion, hardware failure) is also insufficient. This study presents a lightweight Efficient-MCMT optimization scheme and a single-view loss robustness verification system. We incorporate the EfficientNet-B0 as the encoder and decoder, leveraging MBConv depthwise separable convolutions and SE channel attention mechanisms to significantly reduce model complexity while preserving critical target feature extraction. To evaluate robustness, we design systematic experiments simulating "local area occlusion" (partial lens contamination) and "full-view occlusion" (complete camera failure). Extensive experiments on Wildtrack and MultiviewX datasets demonstrate that our approach achieves 60.410% reduction in model parameters and 39.825% reduction in computational complexity. The model retains competitive performance on core metrics (MODA, MOTA, IDF1) compared to state-of-the-art methods, with performance fluctuations within 2% under single-view loss conditions. This work address critical gaps in lightweight in-vehicle adaptation and single-view loss robustness, enabling reliable MCMT deployment in resource-constrained scenarios.