Drone Detection and Tracking in Distorted Surveillance Video

The widespread integration of drones across various industries has led to a surge in production efficiency and innovation. However, this increase in drone usage also brings forth significant threats to security and privacy. In response, this paper aims to develop an efficient tracking system for predicting future drone positions in surveillance footage using LSTM and Bi-LSTM. Surveillance videos from the Drone Detection dataset, which includes challenging attributes such as varying drone sizes, abrupt drone motion, and complex scenes, were utilised. Leveraging this dataset enabled a comprehensive evaluation of the tracking system across various scenarios. Coordinates of the bounding box centre was extracted from selected videos to create CSV files for training and testing. LSTM and Bi-LSTM layers with diverse configurations were employed, and the Bi-LSTM (32), Dropout (0.5) configuration was found to be the best, with the lowest MSE, RMSE, and MAPE of 0.00121, 0.03469, and 4.999%, respectively. This model demonstrated superior performance in accurately predicting the future coordinates of moving objects, validating the effectiveness of the chosen configuration.