Performance Evaluation of Faster R-CNN for On-Road Object Detection on Graphical Processing Unit and Central Processing Unit

摘要

On road object detection is very active research area for autonomous cars driving, pedestrian detection etc. Despite recent momentous enhancements, on road object detection is still a challenge that calls for more accuracy. In this study, we present the implementation of Faster R-CNN training for on road object detection and recognition. We have trained the model with our own dataset categorized into three classes such as car, cycle, pedestrian and test against three different datasets, such as KITTI dataset, video of Beijing road and also tested on our dataset to check the performance of the Faster R-CNN on GPU and CPU. we used this data to utilize Faster R-CNN and to analyze the impact of several factors like training datasets size, pre training model, iteration time, and training methods on the detection results of vehicle and pedestrians. Training the Faster R-CNN model by our own dataset on GPU has took 12 h while CPU took 11 days and 9 h to complete 50000 iterations. The GPU processed the video with a frame rate of 8 fps while CPU processed with 4 fps. The result shows that Faster R-CNN on GPU has higher mean Average Precision than CPU.