Modified Cascade RCNN Based on Contextual Information for Vehicle Detection

摘要

In the process of traditional vehicle detection, there are some problems such as thefault detection and missing detection for small objects and shielded objects. Therefore,we propose a modified Cascade region-based convolutional neural network(RCNN) based on contextual information for vehicle detection. Firstly, the featurepyramid is improved to integrate the shallow information into the deep networklayer by layer to enhance the features of small objects and occlusion objects. In here,we introduce the predictive optimization module and combine the context informationof the region of interest (ROI), which makes the feature information havestronger robustness. Meanwhile, the multi-scale and multi-stage prediction is realizedthrough the multi-threshold prediction network of internal cascade. Under thepremise that the network parameters are basically unchanged, the accuracy rate isimproved. Secondly, the multi-branch dilated convolution is introduced to reduce thefeature loss during the down-sampling process. Finally, the region of interest andcontext information are fused to enhance the object feature expression. Experimentalresults show that the new Cascade RCNN method can better detect small andshielded vehicles compared with other state-of-the-art vehicle detection methods.