Survey on Image Based Object Detectors

Abstract

Sensor fusion-based detector utilizes camera sensors to solve the problem of recognizing objects and their classifications accurately. This has been proven to increase accuracy compared to single sensor detectors and can significantly help with the 3D tracking of vehicles in the sensor system’s area of interest. Even at a distance, where no lidar points are available from the target, a high-resolution camera-based detector can easily detect and classify vehicles. There is a variety of real-time capable 2D object detector convolutional neural networks, some of which are open source. This survey compiles a list of these algorithms, comparing them by precision scores on well-known datasets, and based on experimental evaluation completed on camera images taken on the ZalaZONE test-track to evaluate the distances at which the detectors first perceived the test vehicles. Additionally, inference times are also compared.