Human Detection and Tracking for Autonomous Human-following Quadcopter

In this project, a quadcopter which can autonomously detect and track a person using Deep Learning algorithm and correlation filter tracking technique on Raspberry Pi is presented. There are two major tasks for autonomous detection and tracking procedure executed by the quadcopter: First, a Deep Learning object detection CNN-based model named MobileNetv2-SSDLite was applied to detect a person. This model is proved to be efficient and fast for embedded system and mobile devices. The MobileNetv2-SSDLite model was originally trained on Common object in context dataset (COCO), we kept the weights in the Feature extractor layers (MobileNetsv2) and fine-tuned the detection layers (SSDLite). The purpose of this re-training task was to specialize the model on human detection and the evaluation after training were 98.6% for mAP@[0.5]IoU and 93.6% for mAP@[0.75]IoU. We used a vision-based tracking method for human tracking task named MOSSE to work along with the detection task. This technique uses the bounding box's coordinates collected from the detection task as the initial learning sample and consecutively updated new person's coordinate by online learning method. MOSSE tracking algorithm can also help with small occlusion due to the online-learning ability. Usually, running a Deep learning model like Object detection costs a lot of computational power, especially for an embedded computer like a Raspberry Pi. However, our work keeps the processing speed of the whole system fast enough for real-time implementation by combining an expensive Deep learning model with an inexpensive image-processing based tracking technique. Our proposed method can achieve 3 to 4 FPS on Raspberry Pi which is faster than 0.63 FPS of the detection algorithm. The mentioned results in this article were carried out by testing in a real-time environment with a self - developed quadcopter model.