A modified faster region-based convolutional neural network approach for improved vehicle detection performance

Presently available algorithms employed for vehicle detection exhibit three main disadvantages: slow detection speed, poor small objects detection, and low detection precision. To solve above problems, the present work proposes a vehicle detection approach employing a modified faster region-based convolutional neural network (R-CNN). Firstly, this approach introduces a deep CNN-based on VGG-16 and inception architecture, and adds a set of convolutional kernels with a 1 × 1 size, called a deep convolutional network (DCN). Then, an accurate vehicle region network (AVRN) and a vehicle attribute learning network (VALN) are designed. The AVRN accurately generates vehicle-like regions in real time, and the VALN detects the corresponding classifications and locations of vehicle-like regions. To improve the detection precision, we introduce corresponding loss functions for the AVRN and VALN. The calculation speed is increased by alternately optimizing and jointly training the AVRN and VALN. Experimental results demonstrate that the modified faster R-CNN approach improves significantly vehicle detection performance relative to existing algorithms, where, compared to the standard state-of-the-art faster R-CNN vehicle detection approach, the mean average precision of the test results obtained by the modified approach is increased by 11% and the detection time is reduced by one-third.