A Multitask Cascaded Convolutional Neural Network Based on Full Frame Histogram Equalization for Vehicle Detection

摘要

In order to improve the accuracy of vehicle detection, a Multitask Cascaded Convolutional Neural Network(MC-CNN) based on Full Frame Histogram Equalization (FFHE) algorithm is proposed. FFHE is used to enhance the image, which can solve the problem of unclear image, low contrast, low overall gray value and uneven illumination when the imaging condition is not ideal. MC-CNN locates and classifies the object in the image. The algorithm first performs FFHE operation on the test datasets. Then the region image containing a single vehicle is obtained by the object location algorithm based on Faster R-CNN. Next, the six classification(bus, microbus, minivan, sedan, SUV, and truck) results of each vehicle are obtained through the object classification algorithm based on CNN. To verify the performance of the algorithm, MC-CNN was trained on BIT-Vehicle and tested on SYIT2018-Vehicle. Compared with Faster R-CNN, the average classification accuracy of our algorithm was improved by 11.5% on SYIT2018-Vehicle dataset.