An Improved YOLOv2 for Vehicle Detection

Abstract

Vehicle detection is one of the important applications of object detection in intelligent transportation systems. It aims to extract specific vehicle-type information from pictures or videos containing vehicles. To solve the problems of existing vehicle detection, such as the lack of vehicle-type recognition, low detection accuracy, and slow speed, a new vehicle detection model YOLOv2_Vehicle based on YOLOv2 is proposed in this paper. The k-means++ clustering algorithm was used to cluster the vehicle bounding boxes on the training dataset, and six anchor boxes with different sizes were selected. Considering that the different scales of the vehicles may influence the vehicle detection model, normalization was applied to improve the loss calculation method for length and width of bounding boxes. To improve the feature extraction ability of the network, the multi-layer feature fusion strategy was adopted, and the repeated convolution layers in high layers were removed. The experimental results on the Beijing Institute of Technology (BIT)-Vehicle validation dataset demonstrated that the mean Average Precision (mAP) could reach 94.78%. The proposed model also showed excellent generalization ability on the CompCars test dataset, where the “vehicle face” is quite different from the training dataset. With the comparison experiments, it was proven that the proposed method is effective for vehicle detection. In addition, with network visualization, the proposed model showed excellent feature extraction ability.

Vehicle Detection and Ranging Using Two Different Focal Length Cameras

“…The other is a convolutional neural network based on the regression, such as YOLO [24], SSD [25], and YOLOv2 [26], which has continuously improved the calculation speed. Sang et al proposed an improved YOLOv2 to improve detection accuracy and speed [27]. However, there are still tradeoffs between accuracy and speed when using convolutional neural networks for real-time vehicle detection.…”

Traffic flow estimation with data from a video surveillance camera

“…Based on this data, Dmitriy Anisimov and Tatiana Khanova [1] have shown that a thoroughly constructed SSD-like detector can run faster than 40 frames per second on the modern CPU while maintaining favourable precision. Another example of good speed-precision trade-off is YOLO v2 architecture [30], which was specialized for vehicle detection via anchors clustering, additional loss normalization, and multi-layer feature fusion strategy.…”

Orientation- and Scale-Invariant Multi-Vehicle Detection and Tracking from Unmanned Aerial Videos

“…They generally report higher accuracy in terms of detecting smaller objects in the large image space. Examples include YOLOv2 [41] with improved loss normalization, anchor-box clustering and multi-layer feature fusion [42], YOLOv3 with additional prediction layers [43], Tiny YOLOv3 with repeated up-sampling and additional passthrough layers [44], and focal loss-based RetinaNet [45]. These approaches are performed on relatively low-resolution videos in order to achieve high detection speed and proper accuracy.…”

“…These approaches are performed on relatively low-resolution videos in order to achieve high detection speed and proper accuracy. Examples include resolution of 960 × 540 pixels at 34 fps with average precision of 43%-79% [38], resolution of 300 × 300 pixels at 20 fps with mean average precision (mAP) of 77% [39], varying resolutions up to 1920 × 1080 pixels with processing time of 0.09 sec per frame (~11 fps) with F1-score of 39% [40], varying resolutions up to 608 × 608 pixels with processing time of 0.038 sec per frame (~26 fps) with mAP of 68% [42], resolution of 960 × 540 pixels at 9 fps with mAP of 85% [43], resolution of 512 × 512 pixels at 75 fps with mAP of 80%-89% [44], resolution of 960 × 540 pixels at 21 fps with mAP of 74% [45]. A comprehensive report on the performance of single-network detectors against two-stage networks can be found in [46].…”