6 Algorithm Flow: SFM-Based OG Mapping on TDA4VM

The TDA4VM SoC, which was designed with SFM and other popular Computer Vision tasks necessary for ADAS and robotics in mind, is unsurprisingly uniquely suited for the SFM algorithm. Figure 6-1 shows the TDA4VM device, and Figure 6-2 shows the algorithm flow in Figure 5-1 mapped on to TDA4VM. The Vision Pre-processing Accelerator, or VPAC, mentioned here contains modules to accelerate different image pre-processing sub tasks such as tone mapping, noise filtering, lens distortion correction, and so forth. The Depth and Motion Processing Accelerator, or DMPAC, contains two modules, a Stereo Disparity Engine, or SDE, to accelerate stereo depth estimation, and a Dense Optical Flow Engine, or DOF engine, to accelerate DOF. For more information on the TDA4VM device, see https://www.ti.com/product/TDA4VM.

Table 6-1 shows how the sub-tasks of the SFM OG mapping algorithm map are tabulated to the different processors in the TDA4VM SoC.

Without acceleration through dedicated HWAs, some of these steps are computationally prohibitive. For example, take the DOF portion of the algorithm. For a 2MP camera input, the C7x DSP consumes 2000 Mega Cycles Per Second (MCPS) to generate 2000 key points. Whereas, the DOF HWA only consumes 394 MCPS. Table 6-2 summarizes the results.

Furthermore, since all sub-tasks in this algorithm can be performed using the DSPs and HWAs, the general-purpose ARM cores and Deep Learning Accelerators in the TDA4VM device can be reserved for other applications.

Next, two example implementations of SFM OG mapping are described. Both of these applications are included in the SDK that accompanies TDA4x devices. The first, is intended for a vehicle operating on the road, and therefore is designed for input data from a camera mounted in the front of the vehicle. The second, is for automatic parking, and therefore uses input from a camera mounted on the side of the vehicle.