Integrated real-time object detection for self-driving vehicles

In this paper, an integrated real-time approach for detecting objects in the captured images of self-driving vehicles has been proposed. For this purpose, we model the object detection as a regression problem on the predicted bounding boxes and their class probabilities. Unlike the previous works, no sliding window or region proposal has been used for predicting objects' bounding boxes. Instead, unified neural network has been performed on the whole image which could predicts the bounding boxes and class probabilities at the same time. Our approach performs at 38 frames per second (fps) while achieving the mAP of 68.2% on KITTI vehicle dataset, which is 41.1% more than 30Hz DPM and 6.4% more than Faster R-CNN. Our approach is about 7 fps slower than YOLO. However, since it has an optimized gripping process, it has less localization error than YOLO, and a performance boost about 5.4% in mAP has been achieved.