Covered Vehicle Detection in Autonomous Driving Based on Faster RCNN
Jiaqi Fan, Tianjiao Huo, Xin Yue Li, Ting Qu, Bingzhao Gao, Hong Chen · 2020
Nowadays, Deep learning algorithm is widely used in image object detection. In the field of autonomous driving, the car senses the surrounding environment with the video taken by the camera. Aiming at the front vehicle occlusion problem, in this paper, we migrate Faster RCNN algorithm in deep learning to our own dataset and we get a neural network model with higher robustness. Through the comparison of alternating optimization training method and approximate joint training method, we adopt approximate joint training method to train the Faster RCNN model. Under this training pattern, we improve the size and proportion of anchors constantly. Finally, we conduct extensive experiments and discover that the average test accuracy gets the highest when the basic sizes of the anchors are 64, 128, 256, 512 and the ratios are 0.5, 0.8, 1, 1.2, 1.5, 1.8, 2. And the average precision in test set is 89.06%, which is 3.2% higher than the model under the traditional hyperparameters settings. The best model we train in this paper can get a higher average precision in the test set. It can almost test every car between the overlap cars and therefore the robustness of this model is very high.