Fig. 1. Simulation of typical road scenes. (a)-(c) Road surfaces with different longitudinal slopes; (d)-(j) object surfaces with different orientations

Fig. 2. Simulated disparity images, disparity gradient images and binarized disparity gradient images. (a) Level road; (b) uphill road; (c) downhill road; (d) left-side plane of the vehicle; (e) right-side plane of the vehicle; (f) forward plane of the vehicle; (g) backward plane of the vehicle; (h) left oblique plane of the vehicle; (i) right oblique plane of the vehicle; (j) vertical plane

Fig. 3. Flowchart of double-path convolutional neural network with late fusion

Fig. 4. Experimental results of KITTI road detection testing dataset. (a)(c) Method of Ref. [4]; (b)(d) proposed RGB-DT-IN algorithm

Fig. 5. Convolutional output visualization of road images. (a) Color and gradient maps of input four channel; (b) output results of convergence layer

Fig. 6. Convolutional output visualization of non-road images. (a) Color and gradient maps of input four channel; (b) output results of convergence layer

Table 1. Evaluation results of different methods on the common dataset%

Table 2. Evaluation results of different methods on the difficult dataset%

Table 3. Processing time of methodss