Reuben Strydom

UAV Guidance: A Stereo-Based Technique for Interception of Stationary or Moving Targets

We present a novel stereo-based method for the interception of a static or moving target, from an Unmanned Aerial Vehicle (UAV). This technique is directly applicable for outdoor applications such as search and rescue, monitoring and surveillance, and complex landing scenarios. Stereo vision is particularly useful for the interception of a moving target as it intrinsically measures the relative position and velocity between the UAV and the person or object. The target position is computed geometrically using its direction, as viewed by the vision system, and the UAV’s stereo height. A Kalman filter computes a reliable estimate of relative position and velocity using the target centroid. The performance of this method is validated by conducting a number of closed-loop interceptions for both static and moving target cases. The mean interception error is found to be 0.01m with a standard deviation of 0.33m in tests with static targets and 0.14m with a standard deviation of 0.24m in tests with moving targets. Our method has been field-tested outdoors and provides results comparable to other vision-based techniques that have been tested under more controlled indoor conditions.

Biologically inspired interception: A comparison of pursuit and constant bearing strategies in the presence of sensorimotor delay

We investigate the effect of sensorimotor delay on the pursuit and interception control strategies in the context of robotics. A first-order lead-lag model is introduced to model interception as observed in nature. This model was studied via extensive simulations that incorporate both a sensorimotor delay and a compensatory feedforward controller, allowing examination of the effects of various sensorimotor delays over a range of target velocity to pursuer velocity ratios. It was found that, with appropriate tuning, both the pursuit and constant bearing strategies operate effectively over a wide range of sensorimotor delays. Additionally, for each strategy, the use of a generalised mean value for this gain provides near-optimal performance, compared to separately optimising the gain for each combination of velocity ratio and delay. Finally, it is demonstrated that the constant bearing approach intercepts the target approximately 20% faster, on average, than the pursuit strategy across varying delays.

Snapshot-based control of UAS hover in outdoor environments

With the emergence of rotorcraft unmanned aerial systems (UAS) in civilian applications, the capability of accurate visual hovering is required in near-ground, GPS-denied, flying operations. Optic flow is commonly used for vision-based guidance and control of UAS, enabling autonomous obstacle avoidance, speed regulation, odometry, etc. This paper presents an optic flow-based method that uses a bio-inspired concept of image matching for the control of drift-free hover in natural environments. Our approach uses a reference snapshot (panoramic image) taken at the desired hover location and, through optic flow measurements, it estimates the rotorcrafts 3D position and velocity relative to that location by matching the current and reference views at each time step. These position and velocity signals are fed to a hover controller. Sensing and control are performed in real-time, at camera frame rate (25Hz) onboard a small-size, custom-built quadrotor, and without additional sensor fusion. Results from outdoor closed-loop flight tests demonstrate robustness against long-term drift, as well as improved hover accuracy when compared to techniques that use frame-to-frame integration of egomotion vectors derived from optic flow.

Biologically-inspired visual stabilization of a rotorcraft UAV in unknown outdoor environments

This paper presents the development and flight testing of a novel and efficient view-based method for the navigation and control of rotorcraft unmanned aerial vehicles (UAVs) in unknown, GPS-denied, outdoor environments. At the core of our system is the Image Coordinates Extrapolation (ICE) algorithm which estimates the UAV 3D position and velocity in real-time by computing the pixel-wise difference between the current view (panoramic image) and a snapshot taken at a reference location (e.g., the hovering position). When combined with a PID flight controller, this simple, but effective algorithm allows a rotorcraft UAV to achieve stable and drift-free hover using image differences only, without the need to track features or to compute optic flow. The performance of our approach is evaluated in closed-loop flight tests on a custom-built quadrotor equipped with an onboard panoramic vision system and flight computer.

TCM: A Vision-based Algorithm for Distinguishing Between Stationary and Moving Objects Irrespective of Depth Contrast from a UAS

This paper describes an airborne vision system that is capable of determining whether an object is moving or stationary in an outdoor environment. The proposed method, coined the Triangle Closure Method (TCM), achieves this goal by computing the aircrafts egomotion and combining it with information about the directions connecting the object and the UAS, and the expansion of the object in the image. TCM discriminates between stationary and moving objects with an accuracy rate of up to 96%. The performance of the method is validated in outdoor field tests by implementation in real-time on a quadrotor UAS. We demonstrate that the performance of TCM is better than that of a traditional background subtraction technique, as well as a method that employs the Epipolar Constraint Method. Unlike background subtraction, TCM does not generate false alarms due to parallax when a stationary object is at a distance other than that of the background. It also prevents false negatives when the object is moving along an epipolar constraint. TCM is a reliable and computationally efficient scheme for detecting moving objects, which provides an additional safety layer for autonomous navigation.

UAS stealth: target pursuit at constant distance using a bio-inspired motion camouflage guidance law

The aim of this study is to derive a guidance law by which an unmanned aerial system(s) (UAS) can pursue a moving target at a constant distance, while concealing its own motion. We derive a closed-form solution for the trajectory of the UAS by imposing two key constraints: (1) the shadower moves in such a way as to be perceived as a stationary object by the shadowee, and (2) the distance between the shadower and shadowee is kept constant. Additionally, the theory presented in this paper considers constraints on the maximum achievable speed and acceleration of the shadower. Our theory is tested through Matlab simulations, which validate the camouflage strategy for both 2D and 3D conditions. Furthermore, experiments using a realistic vision-based implementation are conducted in a virtual environment, where the results demonstrate that even with noisy state information it is possible to remain well camouflaged using the constant distance motion camouflage technique.

WHoG: A weighted HoG-based scheme for the detection of birds and identification of their poses in natural environments

We describe a technique for object detection that uses a combination of global shape descriptors and local point descriptors. Our system is able to represent pose using a global shape descriptor, rather than the commonly used part based representation. This approach considerably reduces computational complexity and achieves a significant performance improvement on an extensive dataset: CUB-200-2011 [31]. Our methodology is valuable for the detection of textured objects that are viewed against background clutter and possess a high degree of articulation and variation of pose, as for example in birds. We demonstrate how high and low frequency gradients can be separated to better deal with the presence of interfering textures or stripes within the body, which is a major problem in the detection of bird-like objects. Furthermore, detection accuracy is improved by integrating appropriately designed scale invariant color features into the algorithm.