Object detection with non-traditional sensors

Today, the navigation in known places is quite easily reached using several solutions as GNSS and visual odometry but when the space is unknown, the navigation is more critical: no reference, no maps. How to do the navigation under this condition? One of the challenges in the Smart Societies this kind of navigation in outdoor and indoor condition, even using UGV (Unmanned Ground Vehicle) system. ICT could have an important role in this domain, considering the competences on the technologies and innovative solutions. These devices can autonomously acquire important data, but they need to use some particular tools and algorithms devoted to investigate and analyze the space. Typically, expensive sensors and robust computer are required to solving the navigation in this complex condition. The goal of this thesis is to test not-conventional sensors, as low cost systems, mass market solution and IR camera for indoor navigation, implementing the algorithms on low cost platform, such as a Raspberry Pi. With a smartphone, photos of an extinguisher were taken, and the training sets were created, one for each algorithm. Once trainings were done, performances of the chosen algorithms were evaluated on sample images taken with four sensors: Smartphone, Official Pi Camera, Longruner Camera, MAPIR Survey3. In term of algorithms, the navigation is partially supported by the Object Detection (OD) which has a significant practical importance and it is used across a variety of fields as: autonomous vehicles, workplace automation and surveillance. The use of autonomous surveillance systems is increasingly common, moreover a UGV could be useful in dangerous situations to identify emergency exits or useful objects such as the extinguisher. There are many models available and, in this thesis, Haar Cascade (HC) and YOLO have been compared. The tools used are: OpenCV to implement HC and Darknet to implement YOLO. Several tests have been done and the workflow can be summarized by these steps: 1. Gathering training data 2. Training the model 3. Prediction 4. mAP (mean Average Precision) evaluation OD suffers from the complexity of creating the training set, because for a good training a huge number of images is needed. The training of both algorithms was done on the images taken with the smartphone, to evaluate whether the sets of images, already present on the web, can be used for the detection on images acquired with different sensors, as IR cameras. Since low cost sensors have been used, the calibration tools of MATLAB and OpenCV have been exploited. The mAP achieved by the algorithms have been evaluated on images with and without distortions. The mAP of HC is lower than the mAP of YOLO, that is not able to identify the extinguisher in the datasets of the MAPIR and the night vision cameras. HC reaches lower mAP compared to the one gained on the smartphone and Official Pi Camera images, but still manages the detection. This greater versatility can be due to the use of grayscale images during training and testing, while YOLO works on RAW images and so has trouble recognizing the object on images different from those used during training. Furthermore, since the application to be obtained should work in real time, it has been tested how descriptors, SIFT and SURF, combined with the RANSAC algorithm, can speed up the detection on videos. However, the results obtained are promising and with the improvement of ICT, the application studied will be more efficient.