Vehicle Object Tracking

In this work we study the object detection and tracking problem for autonomous vehicle navigation by tracking cars. Our objective is to track cars from a video sequence by drawing a bounding box over these objects. We plan to use the Nvidia images dataset to do this task. We seek experiment with methods that do the task at hand using geometric computer vision based feature and object detector, such as Histogram of Oriented Gradients and learning based convolutional neural networks for object detection such as YOLOV3 and YOLOV5 . We observed that the performance of YOLOV5 was the best with the best Mean Average Precision among the three models accounting for occlusions and multiple detections. We study the fusion of the detectors with the trackers using association methods based on intersection over union and Kalman Filters and trackers based on convolutional neural networks such as SORT and DeepSORT. We also study the performance of different combinations of detectors and trackers.