VEHICLES AND TRAFFIC SIGNS OBJECT DETECTION AND CLASSIFICATION USING DEEP NEURAL NETWORK APPROACHES IN REAL-TIME CONDITIONS

This paper deals with a comprehensive project in order to create a system for detecting, recognizing, and classifying objects on the road in real-time conditions, with further use in unmanned vehicles. The evolution consists of several steps. The first is to create a classification model based on SVM for two classes of images recognition (vehicle and road signs in general); the second is an improvement of the constructed model using standard CNN. After gaining high-precision results, we moved on to the third step of development, which is building combined CNN model with YOLOv2, as well as performance cooperation with Faster R-CNN. Description of the third step is mainly discussed in this paper. The model trained and tested on our own data set, received from different countries (most of the video data was recorder in real-time, in Ukraine). Along the course, there were problems with signs detection, the localization on a video fragment and recognition accuracy. This led to the necessity of a thorough analysis of the model and additional algorithms. Finally, experiments demonstrated that our model is competitive with state-of-art models