Pedestrian detection using multiple cameras located in the vehicle

U ovom radu opisane su postojeće metode detekcije objekata u slikama. Da bi se razvila potpuna autonomna vožnja kojoj automobilska industrija stremi, ADAS moraju obavljati funkcije s visokom točnošću i preciznošću. Nakon opisa postojećih metoda za detekciju objekata, u sklopu ovog rada razvijeno je vlastito rješenje za iscrtavanje 2D mape položaja pješaka u okolini vozila. Prvo je razvijen vlastiti slučaj upotrebe unutar VisionSDK razvojnog okruženja, koji se zasnivao na algoritmu za detekciju objekata, koji je dio VisionSDK razvojnog okruženja. Nakon razvoja vlastitog slučaja upotrebe, započet je razvoj vlastitih algoritama za procjenu udaljenosti i za iscrtavanje 2D mape položaja pješaka u okolini vozila. Vlastiti algoritmi su izrađeni u C programskom jeziku, a konačno rješenje je implementirano na ADAS ALPHA ploči. Evaluacijom rješenja je utvrđeno da je algoritam za detekciju pješaka nedovoljno pouzdan, ali da ima zadovoljavajuću razinu odziva. Rezultati procjene udaljenosti su pokazali da detekcija i procjena udaljenosti uvelike ovise o rezoluciji okvira video signala na kojima se detekcija pješaka izvršava. U konačnici, krajnji rezultat iscrtavanja 2D mape položaja pješaka u okolini vozila postiže zadovoljavajuće rezultate na manjim udaljenostima od kamere. Kalibracijom kamera i modificiranjem algoritma za detekciju kao i vlastitih algoritama, ispravnost rada konačnog rješenja mogla bi se povećati.This master’s thesis describes existing object detection methods in images. In order to develop the fully autonomous driving that the automotive industry strives for, ADAS must perform functions with high accuracy and precision. After the description of the existing methods for object detection, as part of this master’s thesis, own solution for drawing a 2D map of the position of pedestrians in the environment of the vehicle was developed. First, own use case within the VisionSDK development environment, which was based on an object detection algorithm, which is part of the VisionSDK development environment, was developed. After developing own use case, own algorithms for estimating the distance and for drawing a 2D map of the position of pedestrians in the vicinity of the vehicle were developed. Own algorithms were developed in C programming language, and the final solution was implemented on the ADAS ALPHA board. The evaluation of the solution revealed that the algorithm for pedestrian detection is insufficiently reliable, but that it has a satisfactory level of response. The results of the distance estimation have shown that the detection and estimation of the distance largely depend on the resolution of the video signal frames on which the pedestrian detection is performed. Ultimately, the end result of drawing a 2D map of pedestrian positions around the vehicle achieves satisfactory results at shorter distances from the camera. By calibrating the cameras and modifying the detection algorithm as well as our own algorithms, the correctness of the operation of the final solution could be increased