Performance Analysis and Comparison of Faster R-CNN, Mask R-CNN and ResNet50 for the Detection and Counting of Vehicles

Traffic congestion is one of the major issues of urban cities. The conventional techniques used usually to control traffic via different types of sensors are less precise and expensive. Intelligent solutions using deep learning algorithms provide promising results in terms of better performance, prompt decision making and cost effectiveness. This article aims at providing an easy, more accurate and less expensive solution for the traffic control issues specifically at the traffic signals. Three deep neural network (DNN) frameworks i.e. Faster R-CNN, Mask R-CNN and ResNet-50 have been implemented and compared for vehicle detection, classification and counting. A dataset of 3200 images of different vehicles is used for the training of the models. The training is carried out on NVIDIA 1060TI 3GB GPU. Trained system is tested on indigenous recorded video data of 8 hours for two routes at a traffic signal. Results demonstrated that the overall detection accuracy of Faster R-CNN and Mask R-CNN is >80%, whereas detection accuracy of ResNet-50 is >75%. The counting accuracies of Faster R-CNN, Mask R-CNN and ResNet-50 are >75%, >70% and >62% respectively. Various error analysis have been carried out to validate the performance of the aforementioned frameworks. Furthermore, a prototype has also been developed by interconnecting the DNN results with Arduino via Serial communication.