TL;DR
This paper tackles the challenge of inaccurate vehicle weighing due to varying motion states and occlusions in dynamic weighing systems. They propose an enhanced vehicle detection and tracking method that combines an improved SSD detector with Gaussian Smoothing Interpolation (GSI) integrated into DeepSORT, leading to more robust motion state extraction and improved weighing accuracy.
What Problem Does This Paper Address?
Dynamic weighing systems are crucial for traffic management, but their accuracy is hampered by real-world complexities like multiple vehicles, varying driving behaviors, and occlusions. Specifically, when vehicles pass through the weighing zone simultaneously, their motion states (speed fluctuations, acceleration/deceleration) significantly impact the accuracy of the weight measurement. Existing multi-target tracking methods often struggle with vehicle occlusion and frequent identity switches, leading to missed detections and inaccurate trajectory data. This paper aims to improve the extraction of precise vehicle motion states to enable more accurate dynamic weighing.
Methods
The core of the proposed solution involves two main components:
- Improved SSD for Vehicle Detection: The authors enhance the Single Shot Detector (SSD) model to address its sensitivity to vehicle occlusion. This involves introducing an "exclusion loss" to the loss function, which helps minimize overlap with non-vehicle objects and improve the localization accuracy of detected vehicles.
- DeepSORT with Gaussian Smoothing Interpolation (GSI): For multi-target tracking, they leverage the DeepSORT algorithm. To achieve smoother and more stable trajectory and speed data, they integrate Gaussian Smoothing Interpolation (GSI) into the DeepSORT framework. GSI uses Gaussian process regression to predict target positions, considering both motion and historical trajectory data, leading to more accurate velocity and acceleration estimations compared to simple linear interpolation.
The paper also details a GPS-based camera calibration method for accurate coordinate transformation and an accuracy compensation strategy based on vehicle motion states.
Datasets
The study utilizes a custom dataset collected from surveillance videos of an integrated highway control system in Xiaolei Village, Tongchuan City, Shaanxi Province. This dataset includes images for vehicle label and axle type detection, with 5621 relevant images processed. The dataset was split into training (9:1 ratio) and testing sets. Additionally, a dataset containing vehicle length, width, height, and speed fluctuation information in the dynamic weighing area was constructed for analysis.
Key Insights and Ideas
- Motion State is Key: The research strongly emphasizes that vehicle motion states (speed, acceleration, deceleration) are critical factors influencing dynamic weighing accuracy, often more so than vehicle classification or size alone.
- Occlusion is a Major Hurdle: The paper highlights the significant impact of vehicle occlusion on detection accuracy in dynamic weighing scenarios and proposes a novel loss function to mitigate this.
- GSI for Smoother Trajectories: Integrating Gaussian Smoothing Interpolation into DeepSORT proves effective in generating more stable and accurate vehicle trajectories and speed data, which is crucial for subsequent analysis and compensation.
- Personalized Compensation: The ultimate goal is to enable personalized weighing calibration by accounting for individual vehicle motion states, leading to a significant reduction in weighing errors.
Key Results
- The improved SSD model, with the added exclusion loss, demonstrates enhanced detection accuracy and robustness, particularly under varying road illumination conditions.
- The integration of GSI into DeepSORT provides smoother and more stable trajectory and speed data.
- The proposed method achieved an accuracy improvement of up to 1.8% in dynamic weighing results after compensation.
- The analysis revealed that vehicle speed and speed fluctuations have a significant impact on weighing accuracy, with deceleration followed by acceleration being particularly problematic.
- The optimized detection algorithm showed improvements in accuracy (93.6% vs. 85.2%), MAP (89.4% vs. 84.3%), and FPS (66.3 vs. 45.7) compared to the original SSD algorithm.
Limitations and Future Work
- Dataset Size and Imbalance: While a custom dataset was created, the paper notes that the amount of data for each vehicle type and axle configuration is insufficient, leading to data imbalance issues that can affect model generalization. Image data augmentation was used to address this.
- Vehicle Length Impact: While vehicle length was identified as a factor, its impact was noted as smaller compared to speed and speed fluctuations.
- Future Work: The paper suggests that further research into how axle length variations affect weighing accuracy is needed. The developed method's ability to enable faster and more accurate extraction of vehicle motion state information in the dynamic weighing area is highlighted as a key contribution for future advancements.
Summary updated