Improving Image Object Recognition Accuracy in V2X Environments Using the CBAM-MobileNet Model

Abstract

Addressing the issue of low object recognition accuracy in real-time intelligent vehicle-to-everything (V2X) environments, where sparse information and severe occlusion of small objects often degrade perception performance, this paper proposes an engineered lightweight detection framework based on YOLOv8s. Reliable object perception in V2X systems is essential for intelligent wireless communication networks and electromagnetic sensing environments, where accurate interpretation of visual information supports cooperative perception and real-time decision-making. The proposed framework replaces the original backbone with a lightweight MobileNetV3-CBAM model, employing depthwise separable convolutions and the CBAM attention mechanism to improve computational efficiency while preserving fine-grained features. The neck adopts weighted bidirectional feature fusion to effectively integrate shallow high-resolution information with deep semantic representations through learnable upsampling and downsampling weights. Furthermore, the detection head is optimized using SIOU loss to enhance localization sensitivity and focal loss to alleviate category imbalance. Experiments conducted on the DAIR-V2X-C dataset demonstrate that the proposed method achieves an mAP@0.5 of 95.5% and an mAP @[0.5:0.95] of 73.4%. For small objects, the mAP@0.5 reaches 88.1%, while under an extreme occlusion level of 0.2, the model still maintains an mAP@0.5 of 80.4%. Meanwhile, the average inference latency remains as low as 13.5 ms, demonstrating its suitability for real-time edge deployment. The proposed framework provides an efficient solution for lightweight perception in V2X systems and offers valuable technical references for intelligent electromagnetic sensing and wireless propagation-aware visual perception applications.