Detecting Flying Objects Using a Single Moving Camera
read more
Citations
Vision Meets Drones: A Challenge
Machine Learning-Based Drone Detection and Classification: State-of-the-Art in Research
Using deep networks for drone detection
Vision Meets Drones: Past, Present and Future
Matthan: Drone Presence Detection by Identifying Physical Signatures in the Drone's RF Communication
References
Random Forests
Gradient-based learning applied to document recognition
Dropout: a simple way to prevent neural networks from overfitting
Histograms of oriented gradients for human detection
An iterative image registration technique with an application to stereo vision
Related Papers (5)
Frequently Asked Questions (14)
Q2. How did the authors improve the performance of the CNN algorithm?
To boost CNN performance, the authors used Local Contrast Normalization (LCN) [46] after every convolutional layer and minimize the Hinge Loss at the final layer of the network, which was shown to be effective [47], [48].
Q3. What is the way to compensate for the motion of the object?
Motion compensation can be seen as a way to make the st-cube invariant from the motion of the aircraft, as it keeps flying object in the center, for all the patches of the st-cube.
Q4. What is the main assumption of the flow dependence?
In addition to optical flow dependence, [31] makes an assumption that camera motion is translational, which is violated in aerial videos.
Q5. How much time does scale adjustment reduce?
Even though adding scale adjustment to motion compensation increases the processing time per st-cube, it reduces the overall computation time by a factor of about 4.
Q6. How do the authors train two different boosted trees regressors?
5One way to predict the translation for an input patch m, is to train two different boosted trees regressors [40] φx(m) and φy(m), one for each 2D direction (horizontal and vertical).
Q7. How can the authors estimate the distance to the camera?
Provided that the camera is calibrated and given the true size of the object, the authors can estimate its distance to the camera, which is critical for collision avoidance purposes.
Q8. How can the authors run the detection to only non-overlapping patches without missing the target?
Applying large translations to the training data allows us to run the detection to only non-overlapping patches without missing the target, as explained at the end of Section 4.3.
Q9. How do the authors train a classifier that takes as input st-cubes?
More specifically, the authors want to train a classifier that takes as input st-cubes such as those depicted by Fig. 4 and returns 1 or -1, depending on the presence or absence of a flying object.
Q10. Why does the algorithm perform poorly in their case?
As shown in Fig. 2, optical flow motion compensation cannot achieve good performance in their case, mostly because the target object is rather small and its appearance can significantly change due to illumination and background changes.
Q11. What are the main categories of approaches for detecting moving objects?
Approaches for detecting moving objects can be classified into three main categories: those that rely on appearance in individualframes, those that rely primarily on motion information across frames, and those that combine the two.
Q12. What is the way to evaluate the motion compensation algorithm?
Another way to evaluate their motion compensation algorithm is to compare the detectors, trained on the data, processed with either HBT-Regression or CNN-Regression methods.
Q13. What does the algorithm do when there is a target object?
As a result, instead of having to test windows centered at every pixel location, the authors only have to check non-overlapping ones because the algorithm will automatically shift their location to center the target object when one is present.
Q14. What is the probability of an object of interest in a st-cube?
During training, the authors write the probability that an st-cube η contains an object of interest (y = 1) or is a part of the background (y = 0) asP (Y = y | η) = e CNN(η)[y]eCNN(η)[0] + eCNN(η)[1] , y = {0, 1} , (3)where CNN(η)[y] denotes the classification score that the network predicts for η as being part of class y and e(·) denotes the exponential function.