|
| Author | Method | Strengths | Limitations |
|
| A. Ammar et al. (2021) | Compare Faster R-CNN and YOLOv3, YOLOv4 with UAV imaging datasets | (i) YOLOv3, YOLOv4 gave higher accuracy than Faster R-CNN | (i) Possible low detecting accuracy with high AUV speed |
| Bin Zuraimi et al. (2021) | Improved YOLOv4 algorithm | (i) Increase the accuracy of vehicle detection systems | (i) Weak performance in case of dense traffic |
| W. Sun et al. (2020) | Optical flow with shadow detection algorithm based on the HSV color space | (i) High accuracy in high shadow settings in daytime | (i) Longer computational time than frame removal method
(ii) Unable to deal with nighttime settings
(iii) Not tested with vehicles in different scales |
| B. Xu al. (2019) | Improved YOLOv3 | (i) Improve the mechanism of calling maps to vehicle detection
(ii) High accuracy | (i) Weak vehicle detection in complex traffic environment |
| L. Chen et al. (2018) | VGG-16 with Image Net dataset, k-means algorithm, feature fusion techniques, fully convolutional architecture | (i) Increased the detection rate of the vehicles with respect to
(a) different scales
(b) vehicles with different appearances and with heavy vehicle occlusion | (i) Not tested in change within illumination environments |
| M. Sheng et al. (2018) | Tested the efficiency of R-CNN & Faster R-CNN model through the increase of training dataset | (i) Able to detect and recognize vehicles from different angles and multiple scenes
(ii) Increase in training datasets results in an increase in vehicle detection rate | (i) Slow detection rate
(ii) Low accuracy rate in vehicle detection
(iiii) Unable to identify vehicles under extreme conditions such as fog and snow |
| X. Li et al. (2018) | YOLO-vocRV | (i) Suitable for the multiple target detection of different traffic densities for vehicles | (i) False detection rate is high
(ii) Requires large training dataset
(iii) Not tested in low-light settings |
| J. Sang et al. (2018) | YOLOv2 model for vehicle detection | (i) Improve the detection of the bounding boxes | (i) Not suitable for datasets that the model is not trained in |
|
