DETECTION OF TRAFFIC OBJECTS FROM VIDEO SURVEILLANCE CAMERAS

Abstract

 The article investigates the problem of detecting traffic objects based on data from video surveillance cameras. The analysis of object detection methods is carried out: from classical approaches using manually created features (SIFT, HOG, LBP) and SVM classifiers, to modern methods based on single-stage convolutional neural networks YOLO and transformers. 
 The aim of the study is to optimize the architecture of the YOLOv8 neural network for recognizing traffic objects, specifically vehicles, from video surveillance cameras. The vehicle class includes cars, buses, and trucks. Particular attention is paid to ensuring high detection quality in difficult conditions, such as night time, adverse weather conditions like rain, fog, or snow, and low image quality. To achieve this, various modifications to the network architecture were explored to balance detection accuracy and computational efficiency, making the model suitable for real-time applications. 
 To evaluate the effectiveness of the proposed approach, a dataset of 1400 images obtained from open sources was created. The images cover different times of day, including dawn, dusk, and nighttime, as well as adverse weather conditions and low-quality images, reflecting real-world scenarios encountered by video surveillance cameras. The dataset was annotated to identify the vehicle class, encompassing cars, buses, and trucks, ensuring robust evaluation under challenging conditions. 
 To optimize the YOLOv8-nano architecture, it was proposed to replace individual C2f blocks with the C3Ghost module in combination with the CBAM attention block. This modification provided a balance between detection quality and computational complexity. As a result of the optimization, the number of parameters was reduced by 26%, and the computational complexity (GFLOPS) was reduced by 10%. At the same time, the detection quality remained comparable to the base model: the average accuracy (mAP@0.50:0.95) decreased slightly from 0.531 to 0.525. 
 The results of the study confirm the promising use of the modified YOLOv8-nano architecture for video surveillance tasks in intelligent transport systems. The proposed approach can be applied in traffic monitoring systems, autonomous vehicle navigation, and other domains requiring fast and accurate video data processing in real time. Future work may focus on further refining the model to handle additional vehicle classes, such as motorcycles and bicycles, to enhance its applicability in diverse scenarios. Additionally, integrating the model into embedded systems for widespread deployment and exploring optimizations for specific camera types or extreme conditions could further improve performance.