Mobile Robot Human Tracking Algorithms
Uploaded by
AFTABMobile Robot Human Tracking Algorithms
Uploaded by
AFTABCommon questions
Powered by AIThe stability-plasticity dilemma in the proposed human tracking algorithm is addressed by using a dynamic object model that evolves over time. This model uses a fixed-size tree of templates, updated online, to integrate new information continually. It reinforces frequently occurring features while allowing the model to adapt to new situations without losing previously learned structures. This balancing act maintains stability in recognizing persistent features while incorporating plasticity to adapt to changes such as pose or lighting variations .
The thesis claims several advancements over existing point-based methods for human tracking, including addressing out-of-plane rotations and pose changes. By developing a SURF-based algorithm using a dynamic object model, the thesis alleviates common point-based constraints like high computational complexity and instability due to variable matching points. Additionally, the introduction of hybrid tracking systems and the innovative application of region growing algorithms to remove background descriptors significantly advance robustness and real-time trackability in dynamic environments .
The use of an affine transformation is critical for projecting points from previous templates because it accommodates changes in scale, rotation, and translation of the target object within the image frame. This geometric transformation ensures that previously gathered key-point data can remain relevant and accurately superimposed on the current image, thus maintaining the object model's integrity despite pose variations. The method allows the tracker to effectively manage continuity and accuracy in tracking as the human body's appearance transforms spatially across frames .
The key challenges in human tracking from a mobile robotic platform include abrupt object motion, changes in appearance pattern such as pose variations, non-rigid object structure, occlusion, and camera motion. The proposed algorithm addresses these challenges by using a novel color-based human detection method that involves color histograms and head and hand creation based on depth of interest to construct the human silhouette. It employs detectors for height, size, and shape to distinguish humans from other objects. For real-time processing, a shape analysis algorithm identifies humans by detecting the two legs apart pattern in the vertical projection histogram. The algorithm also employs an unscented Kalman filter to predict human location and a dynamic object model with SURF descriptors to handle pose changes and mitigate limitations related to insufficient matching points and computational complexity .
The integration of color information enhances SURF-based human tracking by overcoming limitations associated with point-based methods, such as insufficient matching key-points and handling background objects with similar color distributions. The proposed system uses a Markov random field-based graph matching algorithm to detect humans when the SURF-based tracker fails. It also imposes a structure on color blobs to eliminate background objects with similar color distributions, making the algorithm robust and effective in dynamic environments .
The proposed algorithm handles abrupt object motion by employing an unscented Kalman filter to maintain tracking continuity. The Kalman filter predicts the subsequent location of the human in the image frame based on previous states, providing a reliable means to anticipate sudden changes in the object's position. This predictive capability allows the tracking system to adjust dynamically and robustly, even when faced with sudden or unexpected movements, ensuring that the robot can follow the target without losing track .
The hybrid tracker combines SURF-based mean-shift with optical flow to enhance human tracking, offering improved robustness and adaptability. This integration allows for online template selection and addresses out-of-plane rotations—an issue not effectively solved by individual point-based methods. The hybrid approach also ensures more reliable long-term tracking by solving the template update problem and dealing with occlusions by using a Kalman filter-based motion predictor. It provides superior performance compared to solely using SURF-based mean-shift or optical flow trackers .
The thesis proposes a SURF-based algorithm that updates the object model over time to address pose changes due to out-of-plane rotations. The innovative technique involves projecting points from previous templates using an affine transformation and confirming pose changes by examining the aspect ratio of the bounding region formed by these projected points. This approach effectively resolves the challenges of maintaining accurate human tracking despite variations in pose .
A k-d tree-based classifier assists in differentiating occlusion from pose change in human tracking by analyzing and interpreting the distribution of descriptor points. When occlusion occurs, the distribution is likely to be irregular or sparse, whereas pose changes result in more structured changes in point patterns due to the rotational transformation. By classifying these distribution patterns using k-d trees, the algorithm can accurately identify the underlying cause—allowing for targeted adjustments in the tracking process .
A visual servo controller plays a crucial role in implementing real-time human tracking algorithms by translating data from tracking algorithms into motion commands for the mobile robot. It processes translational and rotational velocities, formulated using the kinematic model of the robot and the pin-hole camera model, to ensure that the robot can dynamically and accurately follow tracked humans. The visual servo controller functions as an intermediary that converts visual cues into actionable navigation tasks, thereby linking the perception system with the robot's motor controls .
