An improved deep learning method for flying object detection and recognition
Shailendra S. Aote,Nisha Wankhade,Aniket Pardhi,Nidhi Misra,Harsh Agrawal,Archana Potnurwar
DOI: https://doi.org/10.1007/s11760-023-02703-y
2023-08-22
Abstract:The demand for detecting and classifying unmanned aerial vehicle (UAV) objects like birds, planes, and drones is increasing in various fields such as the military, surveillance, etc. A distant object appears to be point size; hence, it is difficult to classify the far-way objects in an image. There is always a trade-off between correct object detection and a confidence value. It is important to detect and classify the object correctly with a high confidence value. This paper introduces a hybrid model based on the combination of a convolutional neural network (CNN) and long short-term memory (LSTM) to detect and classify a UAV. Initially, we presented a comparative study of different algorithms from the You Only Look Once (YOLO) family. We have also gathered and prepared a dataset of images from various sources like GitHub, the University of California Irvine Machine Learning Repository (UCI), and the International Conference on Computer Vision (ICCV) for experimentation. The proposed CNN-LSTM model extracts spatial characteristics of the input video sequence and the better memory capacity of LSTM provides best-memorized results for object detection. The Bayesian optimization is used for hyper-parameter tuning that makes the results of the proposed hybrid CNN-LSTM model more promising when compared to the other state-of-the-art algorithms like YOLO, R-CNN, faster R-CNN, SGD, and CNN. We have also presented the detection accuracy with varying distances. The proposed model performs best w.r.t. precision, recall, training and validation accuracy, and loss. The processing speed per second (FPS) is also nearly equivalent to faster R-CNN.
engineering, electrical & electronic,imaging science & photographic technology