MULTI-CLASS VEHICLE DETECTION AND CLASSIFICATION FOR TRAFFIC SURVEILLANCE USING YOLOV8 NANO

Abstract

This rapid urbanization has led to an increasing number of vehicles on the roads, which has created a need for automated and intelligent traffic surveillance systems that can detect and classify various vehicle types in real-time. Current computer vision techniques and manual inspection processes do not effectively deal with the complexity, scale and variability of today’s traffic conditions. This paper introduces an end- to-end deep learning solution for multi-class vehicle detection focusing on road traffic images with eight vehicle classes: Car, Auto, Bus, Truck, Light Commercial Vehicle (LCV), Motorcycle, Tractor, and MultiAxle. An extensive data preprocessing pipeline was created that includes image resizing, removal of corrupt images, optimization of compression, removal of duplicate la- bels and verification of the data set. The YOLOv8 framework automatically applied data augmentation in training, such as horizontal flipping, HSV color adjustment, translation, scaling and mosaic augmentation. The model was trained for 25 epochs, with the AdamW optimizer and split into train/validate set at 80:20. The proposed system achieved a final precision of 0.63, recall of 0.69, mAP@50 of 0.67, and mAP@50-95 of 0.44. All three loss components were found to be decreasing uniformly in both the training and validation sets as confirmed by the convergence analysis, there was no overfitting. The results show that the proposed pre-processing methodology and training setup is capable of providing a reliable multi-class vehicle detection which is efficient to be deployed in real world traffic surveillance.