Synthesis of annotated images as dataset for vehicle counting neural networks using semi-supervised learning

• Main Author: Chan, Patrick Matthew J.
• Format: text
• Language: English
• Published: Animo Repository 2022
• Subjects: Computer vision; Vehicle detectors; Supervised learning (Machine learning); Neural networks (Computer science)
• Online Access: https://animorepository.dlsu.edu.ph/etdm_ece/19; https://animorepository.dlsu.edu.ph/cgi/viewcontent.cgi?article=1023&context=etdm_ece
• Institution: De La Salle University

An intelligent vehicle counting camera network has the potential to provide automation, aid, or both, for many processes involved in the development of smart cities. Some example applications include parking slot management and fee collection, criminal car tracking, parking anti-theft, contactless road violator apprehension, etc. The commonly used approach for vehicle counting algorithms is through fully supervised Convolutional Neural Networks (CNN). However, deploying these systems still requires vast amounts of manual data annotation for practically every single camera added to the network. As a result, expanding this intelligent network to be able to cover a wide area ends up being a slow and expensive process. This study proposes a method of integrating innovations in the recently emerging field of semi-supervised learning (SSL), into alleviating this issue for an existing workflow. Due to the core advantage of the SSL paradigm, the proposed approach can significantly reduce time and labor costs by greatly reducing the amount of manually annotated data needed; thus, paving the way for a more commercially viable usage of vehicle counting-based technologies. Using a separate neural network based on CycleGAN, the size of the training dataset for the existing workflow can be augmented in a new way, via a synthesized “training dataset” generated from the already available datasets of previously deployed cameras. Here, the approach is tested by checking the changes in the mean average precision (mAP) values of the detectron2 core of the vehicle counting network, after the addition of the synthetic dataset to its training pool. Upon evaluation on the CATCH-ALL vehicle detection dataset, the proposed method provided an improved object detection performance from an mAP of 71.644 to 77.523. This improvement was achieved, despite both runs starting from COCO pre-trained weights, and including classical augmentation approaches like random flipping and shortest edge resizing.