Research On Lightweight Object Detection And Scene Classification Algorithm Based On Deep Learning

With the development of computer vision,various technologies based on deep learning have been integrated into Advanced Assistant Driving System(ADAS).How to improve the performance of ADAS from the algorithm and system aspects has become one of the current research hotspots.In this thesis,lightweight object detection and scene classification algorithms with applications in ADAS are taken as the major research subjects.With respect to object detection,an improved lightweight algorithm based on YOLOv3-Tiny is proposed.Two improvements are proposed.First,depth separable convolution and inverse residual is incorporated in the feature extraction of the YOLOv3-Tiny backbone to improve the classification performance and reduce the network computation.Secondly,the attention mechanism is adopted in the proposed method to improve the representation capability of the network.Specifically,the attention module infers the attention weights from the feature map by convolution layer along two dimensions of channel and space.Then the attention weights are multiplied with the original feature map to adjust the features.Both improvements are comprehensively evaluated on three datasets.The first is the well-known benchmark dataset of VOC 2007+2012,which is constructed for common-purpose object detection.The second is the public KITTI dataset which is collected from transportation scenario.The third one is a self-collected ADAS aiming dataset.The experimental results clearly demonstrated the effectiveness of the proposed method.Compared with the benchmark method,our method achieved the improvements of 3.9% and2.1% in m AP(mean average precision)on VOC2007 and the ADAS aiming dataset separately.With respect to ADAS scene classification,an ADAS scene image dataset is collected and an improved scene classification algorithm based on convolution neural network is proposed.The ADAS scene dataset contains 74,000 images of various lightings,roads,weathers and time periods with a front-looking camera mounted on front window.For the classification scheme,the convolution neural network with multi-task branches is designed aiming at multi-label classification of the scene data.Besides,the pre-trained models are used to initialize the convolution neural network for better performance.The experimental results on test set show that the proposed method is quite effective with an average accuracy of92.2%.