Counting vehicles types using deep learning algorithm in video surveillance systems

Detection, identification, and automatic counting of vehicles using video surveillance cameras plays an essential role in intelligent transportation management. Despite the progress that researchers have made in these cases, its operational implementation still faces challenges such as various environmental conditions, unbalanced data sets, accuracy, and speed. Therefore, research is fundamental in solving these issues. This research investigates the use of deep convolutional neural networks (CNNs) for vehicle detection, classification, and counting in video surveillance applications. We analyze recent research trends and challenges, examining the suitability of popular datasets like MSCOCO, PASCAL, ImageNet, DAWN, DETRAC, and MIO-TCD for validating proposed algorithms. Our study focuses on overcoming challenges like poor lighting conditions and unbalanced traffic scenarios. We propose a novel approach for counting vehicle types based on the YOLO model. The proposed method leverages the YOLO model for object detection and recognition, the ByteTrack algorithm for tracking vehicles across frames, and the CLAHE pre-processing and the Multiple-Zone-Line approach for addressing video surveillance challenges. Experimental results on three benchmark videos demonstrate the superiority of YOLOv8 over YOLOv9 in our benchmarks. Moreover, the application of CLAHE pre-processing algorithms further enhances performance. This research offers a promising solution for accurate and robust vehicle detection, classification, and counting in real-world video surveillance scenarios.