CNN-Based Hybrid Approach for Comprehensive Vehicle Detection in Complex Environments

Many detection algorithms (like Faster R-CNN or Mask R-CNN) employed for vehicle detection in composite traffic scenarios struggle to achieve real-time processing speeds and detection accuracy. This paper details a CNN-based hybrid approach employing feature extraction of Convolutional Neural Networks (CNNs) with an integrated Region Proposal Network (RPN) for generating high-quality region proposals. The proposed model uses ResNet-50 as the backbone in Faster R-CNN, so that the network benefits from ResNet-50's deep feature extraction capabilities, which help the RPN generate better region proposals and improve the accuracy of object detection. Thus, enhancing accuracy and robustness in real-world composite traffic scenarios. Further, to maintain speed that is required in real-time, the model skips Feature Pyramid Network (FPN) and spatial attention mechanism as they increase the computational load at different scales making approach slower. We carried out our evaluation on KITTI Vision Benchmark Dataset, Cityscapes dataset and custom dataset designed to test the model's robustness with various vehicle types, lighting conditions, and environmental complexities. The evaluation demonstrates that the proposed hybrid model resulted in a mean average precision (mAP) of 88% (in case of motor cycles) while maintaining real-time processing speeds. Along with the real-time processing speed, the model could also consistently handle detection under various scenarios (e.g., mixed traffic, fog, rain, and low-light conditions) with an acceptable precision. The paper contains findings to highlight the potential of this hybrid approach in applications involving vehicle detection accuracy and speed.