量子密钥分发在智能电网网络安全系统中的适用性研究

Vehicle classification and detection has been a field of application for deep learning and image processing which play a very important role in intelligent transport management and AI-assisted driving systems. In this paper, we have presented a vehicle classification and detection system to detect and classify low-speed and high-speed Bangladeshi vehicles. To begin, we have implemented and tested the performance of the 11 pre-trained deep convolutional neural network (CNN) models: YOLOv8 Classify, MobileNetV2, GoogLeNet, AlexNet, ResNet-50, SqueezeNet, VGG19, DenseNet-121, Xception, InceptionV3, and NASNetMobile on the six vehicle classification and detection datasets: BIT-Vehicle, IDD, DhakaAI, Poribohon-BD, Sorokh-Poth, and VTID2. We have found that YOLOv8 Classify, MobileNetV2, and GoogLeNet models outperform other models in categorising vehicle types (e.g., car, truck, bus) in images where the vehicle is already the main subject. Next, we have customised the LabelImg image annotation tool to improve the following features: (a) Changing Label Font & Border, (b) Detecting Incorrect Labels, (c) Abbreviating Label Names, (d) Improving Crosshair & Bounding Box Guide, (e) Adding Progress Information, and (f) Improving File List Panel. We have collected data from real-world highway conditions in Dhaka city and labelled 54,556 objects from 5,460 images based on 16 Bangladeshi on-road vehicle classes. Furthermore, we have built a Bangladeshi native vehicle detection classifier for locating and identifying vehicles within larger scenes, often with multiple objects using YOLOv8 Detect and SSD-Mobilenet V2 models and later deploying this classifier into NVIDIA Jetson Nano Developer Kit (a small and powerful computer). Finally, we have tested the proposed Bangladeshi vehicle detection system with different timing, lighting, and weather conditions in several areas of Dhaka city. The proposed system can detect and classify low-speed and high-speed vehicles with an average 93% detection rate and 98% accuracy, while facing challenges that include issues with image annotation tools like poor label visibility, lack of error checking, and limited guidance, as well as difficulties in setting up the NVIDIA Jetson Nano embedded device for efficient model deployment.