A Novel approach for Detection and Counting of Vehicles using CNN architectures

Recent years the usage of automobiles significantly increased in urban regions. As the usage increased, the traffic jams also became a major issue in day-to-day life. Even governments tried to increase the infrastructure and failed mitigating the traffic issues. In this research, we are proposing a Deep learning approach to find out the density of the vehicle flow in a fully traffic regions. In order to do this project, we are proposing VGG16 a CNN architecture for vehicle recognition (classification) and proposing Yolo for detection of vehicles. Further we employed the OpenCV techniques for 1) Adaptive background subtraction 2) image segmentation 3) vehicle counting 4) vehicle tracking purpose. The primary objectives of this proposed system is: 1) the ability to count vehicles 2) its efficiency, as it produces accurate results. For this we are using a Traffic and Congestions (TRANCOS) dataset which consists of 1244 images of different vehicles, for training and validation purpose and also we gathered an open source traffic surveillance video for live tracking of vehicles and counting. The proposed system achieved 95.77
^%
model accuracy and 98% of ROC curve for both bike and car classes. Further we validated our model using Accuracy, precision, Recall, F1 metrics and achieved 96% accuracy.