Framework for Automated Synthetic Image Generation for Vehicle Detection
Vitória Biz Cecchetti, Bruno José Souza, Roberto Zanetti Freire · 2023
The use of synthetic images has been highlighted within the computer vision field as a more practical option that helps reduce the time and cost spent on vehicle detection tasks in urban environments. This paper presents a framework for automated synthetic image generation, aiming at the detection of vehicles in urban environments. The creation of the 3D scenario is carried out through the Blender software. Some adverse conditions of the scenes were developed, including: rain, fog, occlusion by tree branches and electrical power lines, and low lighting. Then the algorithm for the automated generation of images was developed based on parameters chosen by the user to generate the scene and to generate the annotations of the images. Additionally, a convolutional neural network with YOLOv5 architecture along with a cross-validation technique was applied to carry out vehicle detections on the developed synthetic dataset, on both real and mixed datasets. The proposed framework generated a consistent dataset that can be used in vehicle detection networks, obtaining a mAP 0.5 above 96% and mAP 0.5:0.95 above 82%, also the precision values remained above 92%, and recall above 95%. It was also observed that the combination of synthetic images with real images can improve the performance of the model.