Supervised Object-Specific Distance Estimation from Monocular Images for Autonomous Driving.

Area of Science:

• Computer Vision
• Machine Learning
• Automotive Engineering

Background:

• Accurate distance estimation is crucial for Advanced Driver Assistance Systems (ADAS) like adaptive cruise control and collision avoidance.
• Traditional sensors (radar, lidar) are costly or offer limited object detail compared to image sensors.
• Monocular (single camera) depth estimation presents a cost-effective alternative for automotive sensing.

Purpose of the Study:

• To develop a lightweight convolutional deep learning model for extracting object-specific distance information from monocular images.
• To evaluate the model's performance across various training strategies and structural configurations.
• To compare the proposed model against the established Monodepth2 benchmark.

Main Methods:

• A novel lightweight convolutional neural network architecture was designed for monocular depth estimation.
• Extensive experiments were conducted on the KITTI dataset, evaluating performance on seven road agent categories.
• The model's performance was rigorously compared against Monodepth2 using the average weighted mean absolute error (MAE).

Main Results:

• The proposed lightweight deep learning model demonstrated superior performance in distance estimation.
• The model achieved a 15% improvement over Monodepth2 in average weighted mean absolute error (MAE).
• The model effectively extracts distance information for diverse road agents including persons, vehicles, and cyclists.

Conclusions:

• The developed lightweight convolutional model offers an efficient and accurate solution for monocular depth estimation in ADAS.
• This approach provides a viable, cost-effective alternative to traditional sensors for automotive safety applications.
• The findings suggest significant potential for deep learning-based monocular vision in enhancing vehicle safety and autonomy.