Research On Pedestrian Detection Technology In Blind Zones Of Large Vehicles Based On Improved SSD Algorithm

With the rapid development of social economy,there are more and more cars on the road,especially large trucks.Because of the structural characteristics of "long,wide and high" of the cart itself,and the driving position is on the left side of the vehicle,the driver can not observe the road conditions on the right side of the vehicle in time through the right rear-view mirror.At the same time,when the cart turns on the right,the radius of the inner front wheel is greater than that of the inner rear-view chaos,which means that there is inner wheel difference The existence of wheel difference is also an important reason for the blind area of vehicles.The larger the number of axles and wheelbase,the larger the inner wheel difference and the larger the right turning blind area.The existence of blind area is the main cause of traffic accidents,so the research on pedestrian detection in blind area of large vehicles has important application value.In the actual scene,pedestrian detection has the characteristics of complex background and diversity of pedestrian scale,which makes the detection more difficult than the general target detection.The algorithm is to ensure sufficient accuracy and meet the real-time requirements.Compared with the current excellent deep learning target detection algorithm,SSD algorithm has double advantages in speed and accuracy.However,the network also has the problems of rich texture detail information of shallow features,insufficient semantic information,rich semantic information of deep features,and lack of texture detail information,which leads to the unsatisfactory target detection effect of the algorithm in complex scenes,especially small The target is easy to miss detection,so this paper improves the SSD target detection algorithm,in order to further improve the accuracy and speed of the algorithm,so that it can be better instantiated in the blind area detection of the cart.Specific improvements are as follows:(1)Drawing on the idea of FPN network through feature fusion,from high-level feature map to low-level feature map through deconvolution to do feature fusion in turn to obtain super-feature map,on the basis of which to reconstruct a new multi-scale feature map model structure.The multi-scale fusion combines the rich semantic information of the high level of the feature map and the detailed detailed feature information of the low level,which significantly improves the accuracy of the model.(2)By analyzing the pedestrian proportion characteristics in the actual scene and the convolution receptive field of the convolutional neural network,combined with the characteristics of the Inception structure,an improved Inception Block is proposed to extract different scales in the form of parallel multi-scale asymmetric convolution kernel The characteristics of the design of a convolution kernel that meets the pedestrian ratio reduce theintroduction of parameters and noise,and improve the algorithm detection rate and inference speed.(3)By analyzing the pedestrian width-height ratio distribution in this scene data set,design the default box for pedestrian detection in this scene,reduce the number of default boxes and make it easier to match pedestrians,reduce the extraction of useless frames,improve pedestrian detection Accuracy and reasoning speed.(4)Replacing the 3×3 convolution kernels by using the method of stacking small size convolution kernels with sizes of 1×1,3×3 and 1×1 to reduce the amount of parameters and calculations to achieve the purpose of model compression.