Niu Yifeng

Vision-based behavior for UAV reactive avoidance by using a reinforcement learning method

Reactive behavior is a classic control pattern in both robotic and biological systems. Inspired by this conception, this paper focuses on the mapping from the visual sensor to the actor controller by using a reinforcement learning method. This paper proposes an actor-critic algorithm based on the RBF neural network to achieve the unmanned aerial vehicle (UAV) avoidance ability. A semi-physical experiment is implemented to verify the effectiveness of the proposed algorithm.

An approach to ground target localization for UAVs based on multi-sensor fusion

In this paper, an approach based on multiple airborne sensors fusion to ground target localization for UAVs is proposed, which is the basis of UAVs applications, such as obstacle avoidance, target confirmation, etc. Challenges from accuracy improvement and system disturbance are considered. Sensor information is obtained by a visual sensor, GPS, and IMU modules. Traditional methods are mainly aimed at an upright position for visual sensors. First of all, an online multi-sensor calibration method using GPS is designed, calibrating the orientation of the visual sensor to the UAVs location. Then the target can be detected using Adaboost algorithm based on acquisition information of different visual sensors. And information obtained by real-time GPS, IMU, altimeter, and visual image based on Kalman filter is fused for target localization. Finally a D-S method for estimating the confidence level of target location is designed for perception disturbance. Experiments are implemented on rotor aircrafts and fixed-wing UAVs and the results show that the method, as feasible as it is, can effectively percept disturbance and improve location accuracy of a certain target.

A method of visual sensor employment planning for UAV target detection

Aiming at the wide variety of airborne visual sensor, this paper presents a method of visual sensor employment planning for UAV target detection. Firstly, according to the adaptive range of the image recognition algorithm and the probability of target recognition, the reasonable position of airborne visual sensor is studied. Then, the influence of external environment contrast and visibility for target detection or recognition is analyzed. Finally the model of a UAV sensor is proposed based on the characteristics of Bayesian probability model. The effectiveness and feasibility of this method are indicated by simulation and practice experimental results.

A visual saliency detection algorithm based on channel selecting in transform domains

A transform domain algorithm using channel selecting is proposed for visual saliency in this paper. This algorithm is able to improve the accuracy of the traditional transform domain saliency detection while preserving their advantage of fast processing speed. Based on the image signature algorithm, we select the channels by calculating the correlation values between each channel. Then, the selected channels, instead of all of the channels, are combined for the final saliency map according to the corresponding combination rule which we designed. The experimental comparison between the proposed and original approaches demonstrates that our method is higher in accuracy while maintaining high processing speed.