Real-Time Object Detection and Classification for Autonomous Driving
Seyyed Hamed Naghavi, Hamid Reza Pourreza · 2018
In this paper, a single deep convolutional neural network for real-time detection and classification of on-road objects has been proposed. The resulted network could to be used for implementing a cost-effective and useful system in the domain of self-driving vehicles. Our network has been trained on KITTI Road dataset and could be used to recognize various on-road objects including vehicles, bicyclist, and pedestrians. The final network processes 448×448 input images at 47 frame per second (fps) on a NVIDIA GeForce GTX960 GPU. Our model achieves 78.4% mAP on the KITTI dataset, which is 11.9% higher than traditional YOLO and 5.2% more than SSD300, two of the top real-time object detection systems. Although our system is about 12 fps slower than SSD300, it is still well above the real-time performance.