SAC-RSM: A High-Performance UAV-Side Road Surveillance Model Based on Super-Resolution Assisted Learning

Wenlong Zheng, He Xu, Peng Li, Ruchuan Wang, Xing Shao

Abstract: Efficiently and precisely identifying small items on traffic highways using unmanned aerial vehicle (UAV) platforms with limited resources is a crucial yet challenging job. This research suggests a speedy, accurate, and component-optimized road surveillance model (SAC-RSM) for UAVs. This model addresses slow detection speed, limited detection of small objects, and deployment difficulties. First, we designed a super-resolution-assisted learning branch in the network to balance the model’s detection speed and accuracy. This branch learns the feature representation from low to high resolution. This branch uses multiscale feature fusion in the encoding stage to enhance the feature representation of small objects, thereby enhancing their detection accuracy. Second, to avoid the problem of cross-layer convolution, which results in the loss of fine-grained information and low-learning efficiency, we propose using the convolution-to-space-convolution (CSPC) module in the backbone network to improve model detection’s robustness. Third, to achieve real-time detection, we realized the model using the Huawei Ascend compute architecture for neural networks (CANNs) framework to enable automatic quantization and parallel inference acceleration. Finally, we deployed the accelerated model to the embedded platform Atlas 200I developer kit (DK) A2. Compared to the baseline model, the proposed method shows significant increases in mean average precision (mAP) values for the VisDrone and DroneVehicle data sets, with increases of 17.4% and 9.4%, respectively. The proposed method achieves frames/s (FPS) of 38.3, which is 2.1 times faster than the baseline model, meeting the requirement for high-performance real-time detection in a UAV environment.