A Lightweight End-to-End Framework for Real-Time Vehicle-Ejected Debris Detection on Edge Devices.

Vehicle-ejected debris detection is a practical but insufficiently studied problem in intelligent traffic enforcement. Unlike static road litter, objects thrown from moving vehicles are usually small, irregular, transient, and easily confused with road textures, shadows, lane markings, and light reflections. In current traffic management, such violations still rely heavily on manual video review or offline inspection, while task-specific datasets and edge-deployable detection solutions remain limited. To address this gap, this study constructs a vehicle-ejected debris dataset containing 4328 annotated image samples collected from real road scenarios. The dataset covers urban and suburban roads, daytime and nighttime illumination, near-range and distant small-object cases, and hard negative samples. To meet the coupled requirements of vehicle-mounted small-object detection and edge-side INT8 deployment, this study develops a hardware-aware lightweight detection framework based on YOLOv8m. The original CSPDarknet backbone is replaced with the convolutional variant of MobileNetV4 to reduce feature-extraction cost, while a scale-specific Channel Alignment Module is inserted between the heterogeneous MobileNetV4 backbone and the YOLOv8m PANet neck to preserve multi-scale feature compatibility. The alignment module uses only BPU-friendly convolution, batch normalization, and activation operations, thereby avoiding deployment-unfriendly operators while maintaining compatibility with INT8 quantization and edge acceleration. The trained FP32 model is quantized to INT8 and deployed on the RDK X5 BPU using the Horizon OpenExplorer toolkit. Experimental results and repeated-seed validation show that the proposed model achieves a consistent accuracy-efficiency advantage on the constructed dataset. In a representative run, the proposed model obtains 93.1% mAP