Car detection in UAV images using the SIFT algorithm with local features clustering approach

In the past few decades, urban environments have expanded much larger than before. One of the most important problems in most metropolises and even small cities is the management of the transportation system. An advanced monitoring system of urban vehicles allows for overcoming the traffic problems. With the development of unmanned aerial vehicles (UAVs), continuous and accurate monitoring of urban environments has been provided for the users. In this research, an efficient method is presented that detects the vehicle in the UAV images. The proposed method is effective in terms of computational speed and accuracy.  

   The foundation of the proposed method is based on the characteristics of the local features in the UAV images. The presented method consists of two main stages of training classification model and detecting vehicles. In the first part, local features are extracted and described by the SIFT algorithm. The SIFT algorithm is one of the most powerful algorithms for extracting and describing local features that are used in various photogrammetry and machine vision applications. This algorithm is robust to geometric and radiometric changes of the images. Due to the high dimensions of extracted features from all the training samples, the BOVW (Bag of Visual Word) model has been applied. This model is used to reduce the dimensions of the features and display the images. Simple and efficient computing is one of the significant features of the BOVW model. At this stage, after producing a library of features, the SVM classification model is trained. In the detection part of the algorithm, the images are entered into the algorithm and the local features are extracted in all images by the SIFT algorithm. The BOVW model is often used to display an image patch. In most researches, this model is implemented by applying a search window to the whole image. This type of methods has a higher confidence level in detection, but it is a very time consuming process and increases the volume of the computations. For this purpose, the approach of points clustering and their representation by the BOVW model is proposed. In this method, features that are within a certain range are considered as a cluster. Euclidean distance is used in image space for clustering. Then, the clusters produced by the BOVW model are displayed. Then, a feature vector is constructed for each cluster. The trained SVM is applied to each of the production vectors and each cluster is classified as a vehicle and non-vehicle. If the cluster is detected as a car, the position of the center of that cluster is marked on the image.  

   The proposed method was implemented on 8 images with a number of different car targets. Also, considering the use of the search window approach in many researches, our results were compared with the results obtained by other researchers. The results show that the calculation time of the proposed method is 82 seconds, while the search window method takes 2496 seconds to run. In order to verify the accuracy of the algorithm, two criteria were used. The first criterion is the “Producer's accuracy”, which represents the proportion of correct detections of the vehicle to the entire vehicles existing in the images. This criterion is 75.79% for the proposed method. The second criterion is the “User's accuracy”. This criterion is obtained by dividing the correctly detected samples into the sum of the correctly and incorrectly detected samples. The User's accuracy criterion has been reported to be 59.50%.  

The value of the Producer's accuracy criterion is greater for the search window method which has led to a more accurate detection of vehicles compared to our method.  This is due to the small moving steps of the search window in the entire image. However, the search window method has increased the amount of the time spent on the calculations. The User's accuracy criterion shows that the proposed method has less incorrect detections. The results indicate that our method has a higher degree of reliability. The average of these two criteria indicates the superiority of the proposed method in terms of the accuracy of the calculations. On the other hand, the proposed algorithm has a great advantage in terms of computational speed compared to the search window method.