Deep Neural Networks For Road Object Detection

Compared with manned vehicles,automatic driving vehicles can reduce traffic accidents caused by human factors.Various kinds of sensors are installed on the automatic driving vehicle to sense the surrounding conditions.Among them,the camera is the cheapest,so the use of image processing technology to detect common objects on the road can play an auxiliary role in automatic driving.In recent years,benefiting from the rapid development of convolutional neural networks,modern object detection algorithms have been greatly improved.The object detection algorithm can be abstracted into two steps: Firstly,extract high-dimensional features with powerful representation from the image,and then train a detector to regress the coordinates of object and give the object category.The convolutional neural network is composed of a series of convolution and pooling operations.The pooling process causes downsampling.Therefore,many different resolution feature maps are generated in the convolutional neural network.For these feature maps,the layers near the input hve a higher resolution and a smaller receptive field,and can focus on the local details of the object.The layers near the output have a smaller resolution and are abstracted from the front layer.They have larger receptive field and rich semantic information.Therefore,in this thesis,feature maps of different resolutions are merged by skip connection to form new feature maps that contain local details and rich semantics information.The context information around the object is very important for the object detection.Dilated convolution can aggregate information at different distances from the convolution center by setting different rates.Therefore,in this thesis,dilated convolution of different rates are used in parallel to extract the context information around the object.After that,we uses dilated convolution in parallel and in cascade to make the features generated in the middle be reusable.Cascading and paralleling make parameters more efficient.Pooling is an integral part of convolutional neural networks because it can reduce computation and quickly increase the receptive field.However,the pooling is also accompanied by the loss of information.In order to alleviate this problem,under the condition that does not destroy the pre-trained model,this thesis uses the “Eltw sum” operation similar to the shortcut connection in the residual network to connect the larger feature map with the smaller feature map and makes the smaller features can get the lost information.Automatic technology has high requirements on the real-time performance of the algorithm.Therefore,the thesis uses the three techniques described above to improve the performance of SqueezeDet.Finally,this paper do a large number of experiments on the KITTI dataset and the results indicate a 4~5% boost.The spread time in the GPU can reach 30 fps.