Adversarial camouflage has become an important direction in the research of adversarial attacks on vehicle detection systems due to their multi-view attack characteristics and full-domain surface coverage advantages. Existing methods are mainly based on the differentiable neural rendering framework, which optimizes the adversarial texture using gradient backpropagation through differential derivable mapping from 3D vehicle models to 2D images. However, this method has the following issues: (1) multitask parallel optimization hinders the convergence of the adversarial texture; (2) smooth constraints lead to the loss of local high-frequency adversarial features; (3) unnatural gaps in texture fitting during rendering. In this paper, we propose an efficient adversarial texture optimization method that focuses on critical region adversarial texture generation and enhances visual concealment through IoU-based gradient weight assignment, localized convolutional smooth, and dynamic position adjustment. In addition, we constructed a multi-view dataset under complex weather conditions, containing more than 40,000 high-resolution vehicle images under complex weather such as heavy rain. The experiments show that the AP@0.5 of our method on detectors such as YOLOv3 is reduced to 18.4% on average under heavy rain interference conditions, which is 3% lower than the FCA method. The ablation experiments verify that the loss co-optimization improves the ASR to 75.63% on average, while the rendering compensation leads to a reduction in texture fitting errors. This work provides a new methodology and benchmark system for adversarial robustness assessment of autonomous driving systems in complex environments.