Performance Evaluation of Convolution Neural Network Based Object Detection Model for Bangladeshi Traffic Vehicle Detection

Abstract

AbstractVehicle detection has numerous applications in modern day like smart toll plaza, parking, traffic management, etc. Many Convolution Neural Network (CNN) based object detection models, designed to train specific object detection and test them in real world. However, the challenge is preparing these models for vehicle detection according specific custom datasets and train them with minimal devices such as low GPU, memory, and test those models with embedded devices. Both accuracy and speed are matter for real-time vehicle detection and counting. In general, real-time object detection model based on CNN are 2 types—One stage method (YOLOv3, SSD) and two stage method (Faster-RCNN resnet50, resnet101). Both of the methods have complex network architecture which makes the real-time detection slow. But two stage method is slower than one stage method and in particular YOLO series is faster as it requires less GPU than other models. This paper has shown the performance evaluation of YOLOv3, YOLOv3-tiny, and YOLOv4-tiny in terms of precision, recall, F1-Score, mAP (Mean Average Precision), IoU (Intersection over Union), and Average FPS (Frame per second) for moving traffic vehicle detection and count them using the dataset named ‘Dhaka-AI (Dhaka traffic detection challenge dataset). Experiments show that YOLOv4-tiny performs better compared to other two models in terms of recall, F1-Score, AVG_FPS and mAP.KeywordsDeep learningReal-time object detectionVehicle detectionConvolutional neural network (CNN)YOLOv3YOLOv3-tinyYOLOv4-tiny