Drone Detection with Improved Precision in Traditional Machine Learning and Less Complexity in Single Shot Detectors

This work presents a broad study of drone detection based on a variety of machine-learning methods including traditional and deep-learning techniques. The data sets used are images obtained from sequences of video frames in both RGB and IR formats, filtered and unfiltered. First, traditional machine learning techniques such as SVM and RF were investigated to discover their drawbacks and study their feasibility in drone detection. It was evident that those techniques are not suitable for complex data sets (sets with several non-drone objects and clutter in the background). It was observed that the sliding window size results in a bias toward the selection of the bounding box when using the traditional NMS method. Therefore, to address this issue, a modified NMS is proposed and tested on SVM and RF. SVM and RF with modified NMS managed to achieve a relative improvement of up to 25% based on the evaluation metric. The Deep Learning techniques, on the other hand, showed better detection performance but less improvement when using the proposed NMS method. Since their biggest drawback is complexity, a modified deep learning paradigm was proposed to mitigate the usual complexity associated with deep learning methods. The proposed paradigm uses (Single Shot Detector) SSD and AdderNet filters in an attempt to avoid excessive multiplications in the convolutional layers. To demonstrate our method, the most common deep-learning techniques were comparatively tested to create a baseline for evaluating the proposed SSD/AdderNet. The training and testing of the deep learning models were repeated six times to investigate the consistency of learning in terms of parameters and performance. The proposed model was able to achieve better results with respect to the IR data set compared to its counterpart while reducing the number of multiplications at the convolutional layers by 43.42%. Moreover, and as a result of lower complexity, the proposed SSD/AdderNet showed fewer training and inference times compared to its counterpart.