Research On Human Detection With Very-low Altitude Unmanned Aerial Vehicle

In recent years, with the development of technology of unmanned aerial vehicles and civilian demand, quadrotor unmanned aerial vehicle becomes the research hotspot. Quadrotor unmanned aerial vehicle has the advantages of low cost, simple construction, strong maneuverability, high safety. Its load capacity is strong, capable of carrying aerial photo platform to excute the monitoring missions in near the ground environment such as indoor and tunnel. With the widely use of quadrotor unmanned aerial vehicle which equipped aerial photo platform, human detection are also applied in the aerial image, the victim rescue and other emerging areas.In this thesis, human detection algorithm is studied for the i mage of very-low altitude unmanned aerial vehicle. According to the different application scenarios, the application requirements are devided into human direction is known and human direction is unknown in this thesis, and use different human detection algorithm for different application requirement.According to the view of aerial photo platform affected by the change of vehicle attitude, based on the analysis of modeling a very-low aerial view, this thesis proposed a solution that the different view of ae rial image correction to the fixed standard view. Aims to restore the image to the view which can use for human detection.For the situation of human body is known, such as monitoring missions, use HOG feature which is the most widely applied at present. B efore HOG feature extraction, this algorithm pre-processes the image and uses perspective transform for image correction. Build image pyramids to obtain a series of different size of images and use sliding window detection method to traverse all images. And then extract HOG feature vector from each detection window. Send each feature vector into the classifier which has been trained by the sample and get the classification result. Integrate all classification and get the result of human detection on aerial image.For the situation of human body is unknown, such as victim rescue, use RGTHOG detection method which was proposed in this thesis. RGTHOG detection method could identify human with variable rotation angle. Firstly, Radial Gradient Transform method was adopted to obtain the rotation-invariance gradient. Then, adopting the method that blocks were overlapped, a plurality of descriptors which contain rotation angle information were obtained and connected linearly into a descriptor group with rotation invariance feature, according to the descriptors' rotation angle. Finally, the human detection algorithm was conducted with the support of a two-level cascaded classifier based on Support Vector Machine.Finally, test these two algorithms and verify their feas ibility by the simulation experiment. Experimental data indicate that the two algorithms can achieve the desired effect. The algorithm which use the HOG feature has better detection performance while the algorithm which using RGTHOG feature can achieve human detection on variable rotation angle.