Using Deep Learning and Low-Cost RGB and Thermal Cameras to Detect Pedestrians in Aerial Images Captured by Multirotor UAV

The use of Unmanned Aerial Vehicles (UAV) has been increasing over the last few years in many sorts of applications due mainly to the decreasing cost of this technology. One can see the use of the UAV in several civilian applications such as surveillance and search and rescue. Automatic detection of pedestrians in aerial images is a challenging task. The computing vision system must deal with many sources of variability in the aerial images captured with the UAV, e.g., low-resolution images of pedestrians, images captured at distinct angles due to the degrees of freedom that a UAV can move, the camera platform possibly experiencing some instability while the UAV flies, among others. In this work, we created and evaluated different implementations of Pattern Recognition Systems (PRS) aiming at the automatic detection of pedestrians in aerial images captured with multirotor UAV. The main goal is to assess the feasibility and suitability of distinct PRS implementations running on top of low-cost computing platforms, e.g., single-board computers such as the Raspberry Pi or regular laptops without a GPU. For that, we used four machine learning techniques in the feature extraction and classification steps, namely Haar cascade, LBP cascade, HOG + SVM and Convolutional Neural Networks (CNN). In order to improve the system performance (especially the processing time) and also to decrease the rate of false alarms, we applied the Saliency Map (SM) and Thermal Image Processing (TIP) within the segmentation and detection steps of the PRS. The classification results show the CNN to be the best technique with 99.7% accuracy, followed by HOG + SVM with 92.3%. In situations of partial occlusion, the CNN showed 71.1% sensitivity, which can be considered a good result in comparison with the current state-of-the-art, since part of the original image data is missing. As demonstrated in the experiments, by combining TIP with CNN, the PRS can process more than two frames per second (fps), whereas the PRS that combines TIP with HOG + SVM was able to process 100 fps. It is important to mention that our experiments show that a trade-off analysis must be performed during the design of a pedestrian detection PRS. The faster implementations lead to a decrease in the PRS accuracy. For instance, by using HOG + SVM with TIP, the PRS presented the best performance results, but the obtained accuracy was 35 percentage points lower than the CNN. The obtained results indicate that the best detection technique (i.e., the CNN) requires more computational resources to decrease the PRS computation time. Therefore, this work shows and discusses the pros/cons of each technique and trade-off situations, and hence, one can use such an analysis to improve and tailor the design of a PRS to detect pedestrians in aerial images