Enhancing Multi-class Vehicle Counting for Night-time Scene Videos using AU-GAN
Posathorn Ploykaew, Kanokphan Lertniphonphan, Punnarai Siricharoen · 2024
Urban traffic congestion has significantly increased necessitating effective management by relevant authorities due to the high density of traffic throughout various periods of the day. Due to the low visibility of vehicles in nighttime conditions, it poses a significant challenge for the detection and classification of vehicle types. Moreover, the labeled dataset of traffic images during nighttime for training the detection model is limited compared to daytime data. Consequently, the model's learning capabilities during nighttime are inefficient. To enhance the detection performance during nighttime and achieve high efficiency, it becomes imperative to augment the dataset for learning. Image transfer model to generate synthetic traffic images during nighttime from a daytime dataset. Additionally, improved detection performance positively impacts tracking and counting efficiency. This research presents a vehicle counting system based on traffic surveillance videos during nighttime. Data synthesis involves augmenting data quantities through image synthesis using an image transfer model, along with adjusting image lighting conditions. The system comprises interconnected subsystems: object detection and classification, and object tracking. The object detection subsystem utilizes the YOLOX model for detecting and classifying vehicles. Due to the modifications in YOLOX, which include transitioning to an anchor-free approach and separating the head layer's functions to include both classification and regression patterns, Additionally, the object tracking subsystem employs the ByteTrack model. ByteTrack excels at tracking objects with confidence scores in detection, both high and low. Which enhanced the overall performance when integrated with the preprocessed data. The system's effectiveness is measured by evaluating the mean average precision (mAP), resulting in a high accuracy of 79.7%. Counting efficiency with synthetic data exhibits superior performance with a mean absolute percentage error (MAPE) of 27%. This improvement is attributed to the increased data diversity, which reduced MAPE by 41% of the baseline counting system error.