UAS Visual-Based Human Detection (Drone the Surveillance)

Lim, Jun Hao (2022) UAS Visual-Based Human Detection (Drone the Surveillance). Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text LIM_JUN_HAO_Thesis.pdf Restricted to Registered users only Download (30MB)

Abstract

Nowadays, search and rescue operations are usually facilitated by drones that are capable of visually detect humans in remote areas. The vision system of the drone is generally built on the convolutional neural network for its capability to extract the features of input images or videos to the finest extent. In spite of this, the performance of the vision detection system of the drone requires consistent improvement as the state-of-the-art models experience extreme difficulty in detecting small objects, which are commonly found in the dataset captured by the drone. Therefore, an improved version of the selected state-of-the-art CNN models will be proposed and compared with those other state-of-the-art models as benchmark based on the training and testing results given the same dataset in terms of AP. The objectives of this project were to analyse the suitable type of convolutional neural network model for human detection in SAR operation, improve the accuracy of the selected state-of-the-art model and simulate a simple SAR operation by capturing video with humans from drone and process the video with both state-of-the-art and proposed models. Yolov3 was selected as the base model and seven models based on it, namely Yolov3_SPP, Yolov3_4l, Yolov3_5l, Yolov3_4lSPP, Yolov3_5lSPP, Yolov3_SPPPAN and Yolov3_SPPPAN5l were proposed. They were trained and tested on Kaggle and inferenced on Colab. A drone was built from kits to perform SAR simulation, where it captured videos of humans to the laptop. The test results revealed that the Yolov3_SPPPAN achieved 8.2% higher FPS than Yolov4 with a marginal reduction on AP, or 4.6% lower. On the other hand, Yolov3_SPPPAN5l obtained 13.73% of AP higher than that of Yolov4, but with reduction of 12.81% of FPS. The conclusion for this project was that Yolov3_SPPPAN was the optimal model for the human detection although it had lower AP than Yolov4 despite Yolov3_SPPPAN5l had higher AP than both of them.