Traditional drone detection technologies are based on radio frequencies, infrared cameras and radar. These technologies require powerful computers and specialized hardware for effective operation. In this work, we focus on real-time drone detection on embedded devices using convolutional neural networks. In their basic form, these networks do not take into account the temporal information available in videos. We propose a dynamic framing method that combines optimised convolutional neural networks and temporal information to achieve better detection performance. The method functions by initially suggesting drone candidates and then performing detection on cropped images around these candidates. We compare the performance of our method against traditional implementations of optimised neural networks. Dynamic framing achieves an average precision of 0.692 for IoU thresholds between 0.05 and 0.5, with an average time of 42.8 milliseconds per video frame. This represents an improvement of more than 1 % compared to the best performing neural network tested. Moreover, at an IoU threshold of 0.25, the proposed method achieves a sensitivity of 0.756, which is 7 % higher than the best neural network. Dynamic framing is not limited to drone detection and can be generalized to real-time detection of other small objects.
|
