Visual tracking of ground and air targets

Object tracking is one of the most important components in a wide range of applications in machine vision, such as building surveillance systems for unmanned systems, human computer interaction for control unmanned vehicles, tracking and object recognition, tracking and landing runways, fire detection, object tracking enemy. Systems based on various sensors are created for control the unmanned aircraft. Technologies of visual control are the most promising. Systems visual control uses video cameras installed on board unmanned aerial vehicles as sensors. Nowadays the camera is the most attractive sensor due to its relative low cost, compact size and easy replacement of the equipment in case of breakdown. An integral part of the control system of unmanned aerial vehicle tracking system is tracking objects. Visual tracking - this is one of the most active areas of research in computer vision. Despite the considerable progress that has been made in recent years, the problem still remains not solvable. We have proposed a method for tracking unmanned aerial vehicles which is trained in the online mode, by means of developed scheme model updates. The ability to retrain the system with further another object tracking is not possible, that improves the accuracy of proposed system, even in dark places. Consider the description of our proposed method of tracking objects in unmanned aerial vehicles. Schema description of the proposed tracking method The proposed method of visual tracking of objects consists of several parts: - representation scheme; - search mechanism; - model update Consider each of the stages separately: 1. Representation scheme Histogram local sensitivity used for the objects represent. Histograms of local sensitivity are invariant to changes in illumination and show good results in the problem of tracking and object detection. 2. Search mechanism Two methods respond for search mechanism: - location of the object based on Part Based Detector (PBD); - find offset of the object in the search area; Localization of objects based on the PBD is required when initializing objects tracking, at the time of loss of the object of the search area and the reference frames to improve the tracking reliability. Search object offset based on the knowledge of the object placement in the previous frame, the histogram of local knowledge sensitivity preceding frame object tracking. Model of a random forest has been chosen for efficient search and matching object from the previous frame to the current frame. Our proposed method is based on the union (addition) of two methods. Due to it the reliability significantly increases. 3. Model update The object is represented by a local sensitivities histogram. Tracking features varies in long time that leads to significantly deterioration of objects visual tracking. Therefore, a method of updating the information of the local sensitivity of the histogram is proposed to improve the reliability: 1. Find all the various regions of the histogram between the template object and the histogram of the object at the current frame. 2. If the number of differ regions is smaller T1 (lower threshold is chosen experimentally), then go to item 7. 3. If the number of differ regions is more T2 (upper threshold is chosen experimentally), then the object is considered lost and run PBD. 4. If the number of differ regions is over T1 and less T2 then run the update method of the histogram information 5. Nearest clusters computed from the training sample in the histogram on the current frame to update the histogram. Random forest model is selected as a classifier model. 6. Updating the template histogram is performed using regions derived from a histogram of the current frame and the regions selected from the nearest cluster on the previous step. 7. Get a new frame. Due to the training sample clustering proposed tracker cannot retrained and track an object of another class. In multiple tracking crossing paths problems turns up. Our method uses an optimization algorithm solved the problem assignments (Hungarian algorithm) to solve this problem. This algorithm makes it possible to separate the two close trajectories. Our work present the results of our method testing in comparison with the method of Track Learn Detect (TLD) on five standard video sequences. Experimental results show that the proposed method provides good results in accuracy and robust of the tracking in comparison to state-of-the- art methods.