Range Analysis for Effective Object Detection Algorithms for Camera in Autonomous Vehicles

An autonomous vehicle must be capable of precise detection of objects that are present on the road for safe operations in all scenarios. Computer vision techniques, which inspect captured image attributes to identify the cluster of pixels pertaining to a specific object, are employed for these purposes. This necessitates a pixel-level understanding of the working of the cameras and algorithms employed. This study endeavors to experimentally determine the maximum range (i.e., the maximum distance from the camera pinhole to the object) of any available general RGB camera, up to which the employed algorithm can accurately identify the object of a fixed dimension (in this case, a pedestrian), keeping in mind that this range also depends on the used algorithm and the training data. This dependence is due to the requirement of object mapping on the sensor plane for successful identification. The calculations of perspective view with intrinsic camera parameters are used to correlate the observations with theory for predictive capabilities. Furthermore, this study compares various object detection algorithms trained on the COCO dataset and fine-tuned on TIAND datasets to converge on the best one for the specified use cases. Also, experimental results show significant improvement in the fine-tuned models compared to existing ones.