Real Time Pedestrian and Object Detection and Tracking-based Deep Learning. Application to Drone Visual Tracking

This work aims to show the new approaches in embedded vision dedicated to object detection and tracking for drone visual control. Object/Pedestrian detection has been carried out through two methods: 1. Classical image processing approach through improved Histogram Oriented Gradient (HOG) and Deformable Part Model (DPM) based detection and pattern recognition methods. In this step, we present our improved HOG/DPM approach allowing the detection of a target object in real time. The developed approach allows us not only to detect the object (pedestrian) but also to estimates the distance between the target and the drone. 2. Object/Pedestrian detection-based Deep Learning approach. The target position estimation has been carried out within image analysis. After this, the system sends instruction to the drone engine in order to correct its position and to track target. For this visual servoing, we have applied our improved HOG approach and implemented two kinds of PID controllers. The platform has been validated under different scenarios by comparing measured data to ground truth data given by the drone GPS. Several tests which were ca1rried out at ESIGELEC car park and Rouen city center validate the developed platform.