The Detection Of Pedestrians And Vehicles From A Drone’s Perspective Based On Edge Computing Devices

Edge computing devices can bring computing resources closer to the data source,reduce network bandwidth consumption,and improve the efficiency and accuracy of object detection systems.In addition,edge computing devices are low-cost,low-energy,and highly convenient,making them ideal for deploying detection models to quickly and accurately detect targets such as pedestrians and vehicles in normal scenes,providing strong support for tasks such as traffic control and disaster relief.However,detecting pedestrians and vehicles from a drone’s perspective poses greater challenges than detecting targets in normal scenes,such as long-term occlusion and severe deformation.Additionally,the computing power of edge computing devices is also a consideration.To address these issues,this article primarily focuses on the following tasks:(1)This work involves creating a custom dataset as well as performing effective data resampling on the Vis Drone2019 dataset.Given the significant class imbalance in the publicly available Vis Drone2019 dataset,various data augmentation techniques such as ACE and dark channel algorithms were used to resample the dataset and achieve a relatively balanced distribution of data.Additionally,a custom dataset with a uniform distribution of target classes was created to train the object detection model and improve its generalization ability.(2)We have developed a lighter and more efficient YOLOv5 s improved model to address the challenges of complex image backgrounds,large variations in target scale,and difficulty in detecting small targets from the perspective of drones.To achieve this,we have incorporated deformable convolutions into the backbone network of the model to increase the receptive field of the feature network.Additionally,we have introduced an attention mechanism to strengthen the correlation between target information in the feature map and important channels,addressing the issue of the original YOLOv5 s network’s lack of preference.To reduce memory overhead,we have used lightweight convolutions,specifically GSConv,in the Neck network of the improved model to reduce the number of network parameters,minimize model redundancy,and improve the detection efficiency of the network.(3)We utilized edge computing devices to perform object detection model inference on images captured by drones and effectively detect pedestrians and vehicles.We conducted optimization research on accelerating the detection model inference using the Tensor RT and RKNN neural network computing libraries.The optimized algorithm models were deployed on Jetson Xavier NX and Rockchip RV1126 edge computing devices.By exploring new solutions and methods for object detection based on different edge computing devices,we obtained the suitability of different edge devices for model deployment and the advantages and disadvantages of model inference acceleration.This paper focuses on the target detection from the perspective of drones on edge computing devices.After considering the trade-off between detection accuracy and speed,a low-cost and high-efficiency vehicle and pedestrian detection system under the perspective of drones on edge computing devices has been developed.Experimental results show that the improved model has a memory size of 14.0MB,which is a 2.8%reduction in memory compared to the baseline yolov5 s model.In terms of detection accuracy,the m AP50 value of the improved model is 0.658,which is a 9.5%improvement over the baseline YOLOv5 s model.In terms of inference speed,the average inference time of the improved model is 4.5ms,which is 6.25% faster than the baseline model.Moreover,the improved model has been successfully deployed on both Nvidia Jetson Xavier NX and Rockchip RV1126 edge computing platforms.