MXT-YOLOv7t: An Efficient Real-Time Object Detection for Autonomous Driving in Mixed Traffic Environments

The perception system is vital for the autonomous navigation of self-driving cars in complex traffic environments. A critical part of this system is object detection, which enables the vehicle to perceive and interpret its surroundings. Object detection in mixed traffic environments becomes particularly problematic due to the increased complexity from diverse interacting objects and other factors such as cluttered, blurred scenes, etc. To address this problem, we present the MXT-Dataset, a novel dataset that captures the complexities of real-world mixed traffic scenarios. We also propose MXT-YOLOv7t, a real-time object detection model designed to efficiently and effectively handle the various challenges in mixed traffic scenarios. This model enhances YOLOv7-tiny P5 by improving the detection rate and reducing inference time. The enhancements include refining the feature extraction network by integrating a lightweight attention mechanism into the ELAN blocks and replacing the activation function in each convolution layer with a sigmoid-weighted linear unit. Then, we evaluate the performance by training and testing the proposed model alongside several other previous YOLO variant models using the same MXT-Dataset. The experimental results show that MXT-YOLOv7t model surpasses YOLOv7-tiny model by up to 6.1%, as measured by mAP@0.5. Furthermore, MXT-YOLOv7t achieves a recall of up to 0.934 and a precision of 0.937, significantly enhancing detection accuracy while maintaining low computational complexity. Therefore, the proposed model effectively performs real-time detection in complex traffic scenarios, offering a lightweight design with high detection accuracy, fast inference times, and minimal computational cost.