Transfer Learning-based Framework for Automatic Vehicle Detection, Recognition and Tracking

An automatic vehicle detection, classification and counting framework that combines successfully deep learning and variational models for solving its computer vision tasks is introduced in this research. Two convolutional neural networks (CNN) are created, trained, validated and tested for detection and recognition tasks. First, a YOLO-v2 detection network is built by modifying a MobileNet-v2 deep network and using it for the high-level feature extraction. Then, the obtained CNN model is trained and validated on a voluminous traffic video dataset. The vehicles whose bounding boxes are detected applying the trained CNN-based detector on the movie frames are then extracted using a level-set based active contour model. They are then recognized with another transfer learning network, created by modifying a GoogLeNet classification model and trained on a large vehicle dataset. Next, the detected and classified vehicles are tracked within the traffic video applying a matching process using a IoU-based metric and bounding box-related conditions.