Robot Object Detection using Intersection over Union with Modified Kalman Filter

The object detection and tracking include locating and identifying objects in an image or vide sequence and handling their identity among various frames. However, the accurate detection and tracking of objects in dynamic environment is complex due to the various object sizes, overlapping that leads to misidentification or track failure. Therefore, this research proposes an Intersection over Union with Modified Kaman Filter (IoU-MKF) for robot object detection. The IoU is effective when used as a measure of the spatial coincidence of the predicted and ground truth bounding boxes for proper localization of objects. Through the use of IoU, the system enhances the detection by eliminating boxes that do not encircle the object appropriately therefore increasing the reliability of detection. Another standard metric denoting IoU measures how close objects detected by the algorithm and actual regions of objects in the frame are to each other. Incorporation with the predictive abilities of a MKF proves this technique efficient in addressing the primary issues of real-time object tracking, spatial accuracy and temporal consistency are achieved. The Dish-20 dataset is preprocessed by selective search approach which reduces the search space and enhances the detection performance. The IoU-MKF achieves 86.45% of Multiple Object Tracking Accuracy (MOTA) which is better than existing approaches.