From: A review on action recognition for accident detection in smart city transportation systems
Authors | Model | Dataset | Fps | Dataset feature | Accessibility | Data collection approach |
---|---|---|---|---|---|---|
Yao et al. [86] | Future object localization | DoTA (detection of traffic anomaly) | 10 | 4,77 videos | Yes—link | Dashcam |
Yu et al. [23] | Deep spatiotemporal graph convolutional network (DSTGCN) | Traffic Accident Data, Taxi GPS Data, Meteorological data | – | Varied weather (Cloudy, Snow, etc.) Road network and PoI (point of Interest) | NA | Sensors Traffic surveillance |
Wang et al. [80] | Spatial temporal graph neural network | PEMS–BAY, META–LA | – | 53,116 videos from 325 sensors, 34,272 videos from 207 sensors | Yes—link | Sensors Traffic surveillance |
Bao et al. [88] | Spatiotemporal GCN (graph convolution network) | CCD (Car Crash Dataset), DAD and A3D | – | 6.35 h, varied weather conditions (snow, day and night), 2.43 h and 3.56 h | Yes—link | Dashcam |
Reddy et al. [26] | Deep Q-learning | Traffic driving data | – | 182 drive sequences | NA | Dashcam |
Fernandez et al. [84] | Two-stream network | PREVENTION dataset | – | 6 h video (80 m around ego–vehicle), 3 radars 2 cameras and 1 LiDAR | Yes—link | LiDAR Radar, DashCam |
Ali et al. [22] | Dynamic deep spatiotemporal neural network (DHSTNet) | TaxiBj, Bike NYC | – | 16 months video recordings | NA | Sensors |
Wang et al. [21] | Spatial–temporal mixed attention graph-based convolution model (STMAG) | Curated Traffic data | – | 2000 videos | No-future release | Dashcam |
Alkandari et al. [89] | Dynamic webster with dynamic cycle time algorithm (DWDC) | – | – | – | NA | Sensors |
Riaz et al. [90] | FWPredNet | KITTI, HTA, D2City | 65 | 600 frames from Kitti, 286 clips, 65 frames and 678 video | Yes—link | Dashcam |
Tang et al. [91] | Mixure LDA and expectation–maximization algorithm | 40–seconds traffic video | 15 | 600*800 frame dimension | NA | Traffic surveillance |
Huang et al. [25] | CNN-traffic incident management (TIM) | Traffic management centers report, IOWA DOT radar sensors | – | 856 crash reports, 29 sensors | Yes—link | Sensors |
Bortnikov et al. [92] | 3D convolutional neural network (CNN) | Video game GTA V, YouTube Car Accident Video | – | 5 h recording | NA | Traffic Surveillance, Dashcam |
Gupta et al. [93] | Time-distributed RNN | DETRAC dataset | 25 | 10 h recording of 376 videos, 99 frames selected from each video | Yes—link | Traffic Surveillance |
Yang et al. [94] | Feature-fused SSD detector | Highway vehicle dataset, ImageNet VID dataset | – | 32,938 vehicle samples, 5354 videos | Yes—link | Traffic Surveillance |
Ijjina et al. [83] | Mask R-CNN | YouTube accident videos | 30 | 20Â s video chunks, varied weather (harsh sunlight, daylight hours) | NA | Traffic Surveillance |
You et al. [85] | Single-stream temporal action proposals (SST) | Causality in traffic accident (CTA) | – | 9.53 h video from 1935 videos, 18 semantic cause and 7 semantic effect labels | Yes—link | Dashcam |
Srinivasan et al. [24] | Detection transformers and random forest classifier (DETR) | CADP | – | 1416 Accident footage | Yes—link | Traffic surveillance |
Hui et al. [95] | Gaussian mixture model (GMM) | – | – | – | NA | Sensors |
Min et al. [87] | Sparse topic model | QMUL junction dataset, AVSS dataset | 25 | 52 min traffic video, 4 s chunks | Yes—link | Traffic surveillance |
Vatti et al. [96] | Accident detection and communication system | – | – | – | NA | Sensors |