A Multi-View Vehicle Detection Method Based on Deep Neural Networks

VD (Vehicle Detection) is one of the main challenges in ADAS. The encountered vehicles can be of different orientations and positions, especially in scenarios such as crossroads. Traditional VD approaches have shown outstanding performance only with vehicle images of limited orientations and positions: head and tail of a vehicle, others such as side views have been shown to be detection challenged due to the diversities. In this paper, DNN (deep neural networks) based approach is proposed to tackle multi-view vehicle detection task, which handles images not only with head-tail of vehicles, but also side views and even partially occluded ones in one unified model. First, a weight lightened DNN has been trained to produce preliminary pixel labeled information. Then, ROI (region of interesting) has been selected based on the yielded information. Finally, another weight lightened DNN-based classifier has been adopted to give accurate identification of the ROIs. The experiments illustrated that our model can accurately search ROIs with quite low missing rate (<2%). The quantity of selected ROI has also been sharply reduced compared with other methods. Also a real-time classification accuracy of ROIs can be up to 95%. Our model employs two weights lightened DNN structures, one to propose ROIs while the other one to make a classification. Experimental results have shown that our model is competitive in both accuracy and performance of multiple position VD detection task.